Trian Fund Management, which Peltz co-founded, said Thursday morning that it “intends to take our case for change directly to shareholders.”

Disney, for its part, suggested the proxy fight stemmed from a personal grudge held by one of Peltz’s allies, former Marvel boss Ike Perlmutter.

Trian said Disney earlier in the day offered to set up a meeting with the entertainment giant’s board, but rejected Trian’s bid to join the board, including the addition of Peltz. Trian did not note in a statement how many seats it plans to seek.

Trian declined to comment beyond its statement.

The news came the morning after Disney added Morgan Stanley CEO James Gorman and former Sky TV boss Jeremy Darroch to its board, a move widely seen as a bid to fend off a potential challenge from Peltz. Former Illumina CEO Francis deSouza will not seek reelection to the board.

“While James Gorman and Sir Jeremy Darroch represent an improvement from the status quo, the addition of these directors will not, in our view, restore investor confidence or address the root cause behind the significant value destruction and missteps that this Board has overseen,” Trian said in a statement.

Disney shares are up about 6% this year, far underperforming the S&P 500. The stock was flat Thursday. Later in the day, the company said it would reinstate its dividend at 30 cents a share for shareholders of record as of Dec. 11, payable Jan. 10. Iger had said earlier this year Disney would bring back the dividend, which it suspended in early 2020 during the first days of the pandemic.

Trian said it owns about $3 billion in Disney stock. The firm has oversight of shares owned by former executive Perlmutter, a critic of Disney chief Bob Iger whom the company fired earlier this year.

Disney fired back Thursday, saying Perlmutter has an ax to grind against Iger. Perlmutter has long complained that Disney had spent too much.

“Mr. Peltz, in partnership with Isaac Perlmutter, a former Disney executive, intends to take its case to shareholders. Mr. Perlmutter owns 78% of the shares that Mr. Peltz claims beneficial ownership of, or more than 25 million of the 33 million shares,” Disney said in a statement.

“This dynamic is relevant to assessing Mr. Peltz and any other nominees he may put forth as directors, as Mr. Perlmutter was terminated from his employment by Disney earlier this year and has voiced his longstanding personal agenda against Disney’s CEO, Robert A. Iger, which may be different than that of all other shareholders,” the company added.

Peltz had earlier pushed for a seat on Disney’s board after Trian took an approximately $800 million stake in Disney. After Iger unveiled a broad restructuring of the company in February, enacting layoffs and cost cuts, Peltz backed off a proxy fight.

But Peltz reignited his push in the lead-up to Disney’s quarterly earnings report earlier this month. The activist investor had been waiting to see what happened with the report to decide whether to make a move, CNBC previously reported.

Iger on Tuesday said he was focused on “building again” and intends to focus efforts on theme parks, ESPN’s upcoming streaming service and improving the studio business.

The reason for this change, on the one hand, is due to different enterprises have their own business focus, and therefore create a large model of the product is also inevitably affected by its objective fact; on the other hand, with the overall development of the large model, its industrialization is also beginning to be more and more mentioned. After all, if it can not be deployed into the touch of the product, and then advanced technology is meaningless. In mid-June, Mercedes-Benz cooperated with Microsoft and became the first car company in the world to integrate ChatGPT into its own in-vehicle voice control system; half a month later, a bicycle company announced that it was equipped with ChatGPT, and Geely recently announced that it would release a full-stack self-developed in-vehicle big model in September. As a bystander in lamenting the light of science fiction into reality at the same time, also can not help but let a person reverie, big model may really be the next wind mouth of automotive intelligence?

I. Big model on board, fasten seat belts

On June 16, Mercedes-Benz, a century-old car company, and Microsoft announced that the two companies are expanding their partnership in AI to integrate ChatGPT into Mercedes-Benz's in-vehicle voice control system. The partnership will allow owners to experience ChatGPT while driving through Microsoft's Azure OpenAI service.The test is scheduled to officially begin on the same day as the 16th, and a total of about 900,000 Mercedes-Benz equipped with MBUX infotainment systems in the U.S. will be able to participate in the test.

The development of technology in recent years shows that the car is becoming more and more into a new type of intelligent terminal. After the emergence of the big model, the relationship between human-car and car-machine is also inevitably affected. Specifically, this impact is mainly reflected in two aspects: 1. the impact on automatic driving. 2. the impact on the intelligent cockpit. The following are separated.

1. The impact of large models on automatic driving. Big models can handle massive data, and at the same time have multi-dimensional analysis capabilities, which can provide more accurate and comprehensive data analysis and prediction capabilities. Keeping the optimization and upgrading of the big model can improve the accuracy and reliability of automatic driving. Like ChatGPT kind of generalized big model, also when the number of participants reaches a certain level and starts to have enough ability, it can fire all over the world. And specifically to the application level, the impact of the big model on automatic driving can be subdivided into the cloud and the vehicle end. In the cloud, car companies can play the innate large parameter capacity advantage of the big model, through the big model to complete most of the data annotation and mining work, cost savings, but also with the help of simulation scenarios to build empowerment. On the vehicle side, the big model can be subdivided into multiple subsidiary sub-models to manage different sub-tasks, saving the reasoning and calculation time on the vehicle side and increasing driving safety. In addition, the cloud to the vehicle end of the perception decision-making integration algorithm is often considered to be the last bottleneck of the automatic driving algorithm, perhaps in the car access to the big model can also be effectively solved, the automatic driving algorithm upgrades may no longer be out of reach.

2. In addition, big models will also have a new impact on the R&D methods and business models of car companies. In terms of R&D methods, due to AI's efficient annotation capabilities are obvious to all, data annotation tasks that used to take a long time can now take only a few hours, the time cost of R&D is greatly compressed, and the big model can deal with multi-modal rich data, such as voice, gesture, vision, etc., which can help car enterprises to deeply improve overall R&D effectiveness, reduce costs and increase efficiency. In terms of business model, the current big model is generally able to speak well, integrated into the vehicle system, people, cars and machines may develop from "employment relationship" to "companion relationship", the big model through the machine learning ability will gradually understand human preferences and habits, and then derive a new The big model through machine learning ability will also gradually understand human preferences and habits, and then derive a brand new commercial value.

   Second, are big models and automobiles a match made in heaven?

   After Mercedes-Benz announced its cooperation with Microsoft, domestic Ideal Auto also released its own large model MindGPT, and Baidu's Wenxin Yiyin was also accessed by many car companies such as Red Flag, Changan, Geely, Lantu, and Zero Run. Therefore, it can be expected that the combination of big models with car companies and cars will become more and more common. From the point of view of the participating car companies, they have their own focus and direction for the development of big models. From a functional point of view, they can be divided into two types: 1. for the communication and dialog field of intelligent cockpit; 2. for intelligent systems such as automatic driving. The former, such as the aforementioned cooperation between Mercedes-Benz and Microsoft, as well as access to the Alibaba AliOS intelligent car operating system of Tongyi Qianqian; the latter, such as Ideal Auto's self-developed MindGPT, which gets rid of the dependence on high-definition maps and makes the car closer to the driving performance of human drivers, and the generative big model of self-driving of the millimetre-end Zhixing's DriveGPT, which helps to solve the problem of cognitive decision-making and ultimately realizes the development of cloud-to-vehicle automated driving. cloud-to-vehicle autonomous driving. In addition, four car companies, Great Wall, Chery, Azure and Xiaopeng, have also registered and applied for several GPT-related trademarks, and it is believed that there will soon be big model-related results coming out.

3. III. Light but no light, still waters run deep

   Cars and big models have strong scientific and technological attributes, and the greatest significance of the combination of the two may lie in the complementarity of each other's strengths and weaknesses, that is, to strengthen the manufacturing and consumer attributes of the big models, while strengthening the electronic and technological attributes of the car. So it seems that the real decision on whether the big model can get into the car also depends on the car companies' own technological strength, which even relates to whether they can occupy the future technological commanding heights.

   In all fairness, the combination of cars and big models is worth looking forward to. A major significance of the big model is to redefine the human-computer interaction and related service ecosystem, on board the car will accelerate the process of electronic consumption of the car application service ecosystem, which will largely change the definition of automobiles, in-vehicle systems, big models and other underlying products. In this regard, Huawei automated driving product department former minister Su turnip words can be said to hit the nail on the head: "in the view of the traditional car manufacturers car is the base, in-vehicle App or other systems are trying to put the computer or AI embedded in this base. Our view is different, the base is the computer, the car is a computer-controlled peripheral. This is the essence of the view is different."

The company has only one product, Pi, which just launched two months ago. if ChatGPT is a human efficiency amplifier, then Pi is a human emotion masseur. Unlike ChatGPT's more tool-oriented setup, Pi's main characteristics are empathy, simplicity, humor and innovation.

Pi's popularity illuminates another path that has been hidden by the light of intelligent AI like ChatGPT - Emotion AI. a track that brings understanding, attention, and care to the user, which may have a larger potential market than cold efficiency-boosting intelligent AI.

Prof. Huang Minlie of Tsinghua University is the researcher who chose to embark on the path of emotion AI in China. In his view, GPT is undoubtedly a paradigm breakthrough, but there is no way for it to meet the needs of different domains, especially in terms of emotion. This path of exploration can be traced back to MIT's Psychotherapy Conversation AI in 1966, which is a much older starting point than current general purpose task assistants like GPT.

Prof. Huang believes that beyond the need for efficiency improvement, the important emotional needs of human beings are now far from being met by AI, and this is a huge and should be explored need. Although AI now only has some basic forms of personality, through specialized data training, AI can already take on some of the work of a primary counselor. In addition, Prof. Huang completely agrees with what Hurley said: once AI understands emotions, it is more likely to control human behavior and even PUA conversation partners, which will lead to more problems of AI abuse. Therefore, in the process of technological exploration, the limitations and governance of AI are also very urgent. But the governance path is clear: it takes only two years to weave a safety net, which is enough time to outrun all predictions of AI extinction. Concerns are legitimate, but panic is not necessary.

The following is the full text of the interview:

Industry needs can't be solved by language modeling alone

Tencent Technology: What was your first encounter with ChatGPT like? Is it a paradigm breakthrough?

Huang Minlie: When ChatGPT first came out, the main feature was its high level of intelligence and its position as a general-purpose task assistant. In the past, when we did similar task assistants, such as ordering food and tickets, which were very traditional tasks, ChatGPT was able to handle various open tasks in one model, and the level of capability really overturned our previous knowledge, as it was able to accomplish many different tasks in the same system at a high level. This can be understood as a paradigm breakthrough. It's very different from the technology path we've taken in the past.

Tencent Technology: many researchers, including Likun Yang, believe that ChatGPT is technically relying on the Transformer model from 2017, and therefore has little innovation. Then how does OpenAI make its model do better than any other model?

Minlie Hwang: ChatGPT's bottom layer is based on Transformer's architecture, so there really isn't much innovation in the model's architecture (there are some new innovations in some recent model designs). In fact, its success is actually an innovation in the form of data plus engineering plus integration at the system level.

Integrated innovation includes, for example, at the data level, OpenAI has actually done a lot of data accumulation and data engineering, as well as high-quality manual collection, labeling, cleaning and so on. Engineering level it is actually also facing some big challenges, in the past we may need to do dozens of GPU cards can be (simpler), but the scale of the model and the data to the extent that it requires thousands of tens of thousands of cards, which will involve a lot of parallel algorithms scheduling and other aspects of the engineering challenges. Finally on the system side, we've actually seen that OpenAI has been iterating on GPT as a product for the last couple of years. In contrast, those models we had before were more or less developed as a kind of project, and after I made the model good and open-sourced it, it had no further iterative update mechanism.

Tencent Technology: As you said, OpenAI as an enterprise will have continuous product iteration. Academia basically develops with a single project or experiment. So how do you, as an expert who has been involved in both academia and the corporate world, see the difference and correlation between OpenAI as an enterprise, and research in academia?

Minlie Huang: The difference is mainly that OpenAI (as a business) has a very strong team of algorithms and engineering, which is the first point. The second point is that it has a lot of computing power resources. Now you see in the academic world to do (artificial intelligence), we are unlikely to have so much computing power resources, and the second is unlikely to have such a large engineering team. So academics are now focusing on some basic problems, such as the big models we see now may produce the problem of hallucinations, the problem of safety, including accurate computation, that is, the model is not able to carry out accurate computation very effectively.

Those engineers and scientists at OpenAI are actually very good at academics, and then they have very good algorithms and engineering skills, so they can do this exceptionally well. I think the future of real AGI must be a product of close cooperation between the top academic institutions and industry.

Tencent Technology: Some time ago one of Google's engineers published an internal meme saying that big language models can have no moat. Including OpenAI doesn't have one, and neither does Google, and everyone will probably surpass them soon. Do you recognize this statement?

Minlie Huang: I think that's also a degree of misunderstanding. From Google's point of view, if you really want to do it seriously, I think it should not be too difficult to catch up with OpenAI, because it has the computing power, data team, and talent. However, if other companies say that it is easy to go beyond, I think there is a feeling of talking on paper. Like saying, the principle of the atomic bomb looks simple, but it is not easy to really do it.

Because in fact here computing power, money, talent, data and other aspects are actually need to spend time to accumulate and precipitation, including these domestic companies now claiming to do China OpenAI. In fact, everyone is catching up, but you can catch up to 80 points 90 points has been very impressive. And people are constantly iterating, constantly progressing, so I think this thing is actually quite complex, it's a system level problem. And not just that there is no innovation in the model structure, that is no moat. In itself, it is a comprehensive strength of the consideration, it is not only the model of the structure of the algorithm of innovation, more may be the arithmetic funds, and then the data, and then the whole engineering level (caused) such a barrier.

Tencent Technology: Do you think that some companies like OpenAI or Google have already established a moat?

Huang Minlie: There is no doubt that OpenAI has its own moat, and it is not easy for others to catch up with him.

For example, the details of GPT4 have not been released, but its multimodal capabilities are still very strong. In addition to OpenAI is also constantly utilizing the data flywheel to continue to progress. China, we also have some companies in the leading stage, but in fact how the future development, who can ultimately win, depends on one is the overall positioning, the other is in this area can continue to invest, how long you can insist. This is basically the logic.

Tencent Technology: Now everyone is using the same model, recently there are some new research can reduce the overall cost of training. Do you think that if Chinese companies are going to break through and catch up with OpenAI, there are some other paths to choose from, rather than taking the exact same route?

Minlie Huang: I think that's a very good question. In fact, now everyone is squeezing the big language modeling track, but in fact I think the future of AGI does not exclude some other routes. A lot of people also question that a big language model like ChatGPT can't actually create anything new at all. So it's very likely that there will be new routes coming out in the future, but people can't see (the specific direction) yet. It's just that right now we're finding that the path of big language modeling might be closer to AGI or a route that's easier to implement. Now to be honest, the other paths it faces is, for example, symbolism, it has a lot of symbolic operations based on symbols, how it can be scaled up in engineering, that's one of the most realistic difficulties.

And now the big language model has been able to not only make it very big and use a lot of data, but also very capable, so I think that's the silver lining that we're seeing at the moment. But in the future I think there's definitely going to be something else, it's possible that in using the big language model as a framework, it's going to fit some other things in, like the semiotics school.

Tencent Technology: before Lu Qi also put forward, do not go back to the knowledge graph, this point of view you agree with?

Huang Minlie: I don't know the background of his sentence. As far as I know, the big model as a knowledge base to do Q&A is quite far from the ability of other traditional methods on benchmark data sets, and some people have done such research. When the current GPT goes to work on some mathematical calculations, it basically just messes up the answers. Because math problems are exact reasoning, you don't say 1+1=3. 1+1=2 (this narrative), it's either 1 (true) or 0 (false), it only has the probability of 0 and 1, it doesn't say the probability between 0 and 1. So symbolic reasoning is very important in many cases.

Tencent Technology: Before Sam Altman also mentioned in the interview, if the general-purpose model is developed very quickly, a lot of tasks it can be done very well. Does it make sense for us to develop verticals?

Minlie Huang: It definitely makes a lot of sense. Doing industry modeling, domain modeling on a pedestal basis, this is actually very necessary. General intelligence model we actually do not need to solve the problem of the final delivery. When you go to an industry to a field, I definitely want to solve some of the real needs of the industry and field, some pain points, which will involve a lot of industry knowledge and rules.

In the process of sinking the big language model, domain and industry-specific training, optimization methods, including how to inject some industry knowledge and rules into it, which is very important to be able to really let it produce value and play a role in the actual business.

For example, in healthcare, there are some cases where you can never be wrong. Here you need some additional algorithms, modular processing. When doing psychological counseling, one of the scenarios we face is to use a large model to correspond to a depressed user, who is very prone to suicide, and may have a mental breakdown while chatting, and then he says he wants to find a rooftop to jump off. At this time you need to immediately detect his state and implement interventions, such as receiving a manual service up. One of the things we do, is to make a very strong classifier to see if he will have suicidal tendencies, as long as in the detection of the relevant tendency will immediately terminate the human-machine dialogue.

In addition, if you are in the financial scenario, you need to be dynamic and real-time information. We are now cooperating with CICC and Ant to do the application of some big models in financial scenarios, which is a breakthrough in how to get dynamic real-time work.

The other thing is that you can't talk nonsense, and at the same time you need to be compliant when recommending stocks and buying funds. This kind of compliance can not be realized through a simple model data-driven approach.

Tencent Technology: What areas do you think are now currently likely to be the first to be changed by the intervention of AI?

Huang Minlie: I think probably the easiest to see are some related to writing, such as the improvement of the efficiency of writing code, as well as like marketing, digital marketing - I can write a marketing copy, and then input a lot of material, and use AIGC to produce the output. Education is also a big scene, for example, now AI-assisted teachers can guide children to be able to better think, better to understand the plot and values in the story.

Other games are also a big scene. Another relatively more difficult to do (field), such as medical, financial. Because it has a lot of dynamic real-time information and knowledge base, we need to better deal with this kind of field and business-related things, which may be a little slower, but will potentially play a big role and value.

Tencent Technology: What are the most important capabilities that companies in the market need to develop vertical domain models?

Huang Minlie: I think one is to have some underlying capabilities, such as pre-training model fine-tuning, reinforcement learning, etc., is still very important.

On the other hand, it means that you have to have an understanding of the industry and the domain. Knowledge of the industry means that I know when and how to embed this industry knowledge and rules into the model. It's not just a matter of taking data and training it, it's a matter of combining it with the underlying algorithms and models.

But there are a lot of details here that are not as easy as they seem. It is not as simple as saying that I take the data over and then train (training) a bit, get a good result, for example, can do 80 points, but in fact, you ultimately away from the delivery of customer demand may be 95 points, this time you this 15 points how to improve? Is the need for some industry experts to participate.

The strike, which began on May 2, 2023, is the largest ever in Hollywood. Close to two months have passed, not only no fire trend, but also intensified - from the initial 11,000 writers involved, gradually involved in a number of actors and actresses, including directors and unions. A number of popular programs and series stop broadcasting, stop filming, the entire U.S. film and television industry into chaos.

On June 21, thousands of writers' union members marched in Pan Pacific Park in the Fairfax, Virginia area, the latest development in the current strike action.

A "pre-publicized" strike action

The strike started simply because the writers' union, the WGA, failed to agree on a new contract with the AMPTP, which represents several major Hollywood studios - the Writers Guild of America, and the Alliance of Motion Picture and Television Producers. Guild of America), the latter is the "Alliance of Motion Picture and Television Producers" (Alliance of Motion Picture and Television Producers).

As early as April this year, 11,500 members of the Writers Guild of America voted overwhelmingly to go on strike if the new contract was not signed. As a result, after the old contract expired on May 1, a new one was not in sight, so the strike began the next day. Because the main participants were mainly Hollywood writers and actors, the strike action was called the "Hollywood Strike".

The WGA demanded a total pay raise of nearly $600 million for the writers, but the AMPTP rejected the demand, and the camp behind the AMPTP includes major companies and platforms such as Amazon, Disney, Universal Pictures, Paramount and Apple.

Hollywood has had many strikes in its history, and each one has had a huge impact on the U.S. film and television industry. The last writers' strike of the same scale occurred in 2007 and lasted for 100 days, causing more than $3 billion in economic losses in California alone, with several shows, including Grey's Anatomy and Desperado, affected, cancelled or delayed.

This strike was no exception. In addition to the suspension of several nightly talk shows such as "The Tonight Show with Jimmy Fallon" and "Weekend Night Live," according to the Washington Post, more than two dozen film and television episodes or projects have been affected, including the ongoing final season of "Stranger Things," sequels to "Avatar" and "Star Wars," the "Game of Thrones" spinoff "Knights of the Seven Kingdoms" and more. [1] Because of the lack of writers, these episodes could not continue to be filmed, and some nightly shows were replaced with reruns of old shows for the time being.

Due to the spillover effect of the strike, film and television-related industries were also affected. People working in the industry, including prop production companies, equipment companies, transportation drivers, and even catering staff on set, may lose their jobs.

The writers' strike of 2007 has contributed to the rise of reality TV shows that rely less on scripts, such as the worldwide hit "Keeping Up with the Kardashians," which was a product of the strike.

In addition to writers, Hollywood actors and directors are negotiating a new round of contracts with the AMPTP. SAG-AFTRA, the union representing 160,000 film and television actors, is negotiating with the AMPTP over a contract that is set to expire on June 30. Union members voted overwhelmingly (98 percent) to authorize union leaders to also strike, as the writers did, if a new contract is not reached by June 30.

The actors' strike will lead to a broader shutdown in Hollywood, and the two sides have now been in contract negotiations since June 7. The Directors Guild of America (DGA), on the other hand, has reached an agreement with AMPTP that includes salary increases (12.5% over 3 years) and benefits packages, increased streaming content splits, and prevention of AI abuse, among other things.

It is worth noting that the DGA was the first to negotiate a contract during the strike action in 2007, and the writers were forced to agree to similar terms and hastily end the strike.

But judging from the current direction, the drama won't repeat itself, and this time the writers don't seem to be planning to compromise.

In the age of streaming, a new labor conflict

According to a previous report released by the WGA, the median weekly salary for writer-producers has fallen 23 percent over the past decade. Meanwhile, nearly half (49 percent) of U.S. writers are receiving minimum wage, and their pay rates have risen by 16 percent from just 10 years ago. [2]

This sounds mind-boggling. After all, the past decade has also been a decade of rapid growth and takedown of streaming platforms. Netflix has hit one statistical miracle after another and taken over half of Hollywood, with a huge number of series being developed, filmed, and aired, generating huge profits, and correspondingly, production investment is rising. However, the income received by writers is quite small, and has even been shrinking.

This has to mention the impact of the streaming platform on the traditional production model, and the resulting decline in income for writers and other issues.

In the film and TV production industry, the traditional model is that writers are hired to write scripts, then receive an advance and a final payment when the show is rebroadcast. A show would employ about 7-12 writers who would spend about 20 weeks in the writer's room to write a full script.

But in the age of streaming, the writer's room has become a mini room. Producers often ask writers to complete an outline, or a script for a pilot episode, first. The subsequent production of the script may be irrelevant to the writer - either the whole project will be stopped due to the poor reception of the pilot episode, or the producer will hire a lower-priced junior writer to finish the rest of the script based on the outline. From the producer's perspective, this model is less expensive, but for the writers, it means lower paychecks and no job security.

Even with the opportunity to write an entire season, writers can earn less than ever before. The reason for this is that due to the nature of streaming platforms, episodes tend to be very short and have fewer episodes. Generally speaking, the number of episodes of traditional American dramas is around 20, but streaming episodes will generally be 6-8 episodes. At the same time, scripts are produced at a much faster pace, with writers often working for 6-9 months in the traditional model, while streaming reduces the production time to a few weeks. This means great work pressure on one hand, and on the other hand, it means that writers' basic benefits are not guaranteed. Because the actual hours of work are reduced, and the producers only provide benefits coverage when the writers are working for them.

As screenwriter Ellie Adelson says, the screenwriting industry is going from a stable middle-class life to a very precarious one, almost a zero-work economy.

According to the WGA, the lowest weekly salary for a screenwriter is only $4,546. While that may not sound like a small number, it's important to know that writers don't receive a regular salary and are only available for a very limited period of time during the year," said Chris Keyser, co-chair of the WGA negotiating committee. us to the brink of survival." [3]

The WGA is also pushing for improvements in other areas, such as fighting for higher residual pay and calling for an industry standard for the number of writers per show. To prevent producers from hiring writers from outside the union to write scripts to break the strike, the WGA said writers who worked for the producers during this time would be ineligible to enter the union.

It's worth noting that many writers are also threatening to dramatize already-written episodes if a new deal is delayed.

AI-generated content that became the focus of the strike

In addition to the impact brought by streaming platforms, AI, especially AIGC (Artificial Intelligence Generated Content), has become the core of this conflict.

ChatGPT, which has been gaining attention since early 2023, has affected Hollywood and the entire film and television industry. Marvel's latest airing of the film and TV series "Secret Invasion" has already used AI in the production process to generate the opening credits and has been subject to controversy.

For screenwriters, the already meager benefits can hardly withstand the impact of AI employment replacement. At the same time, the AI trains a library of existing scripts that make up a large portion of it. This is equivalent to the AI stealing the fruits of a screenwriter's labor without any payment. Therefore, the various demands against AI have become the core of what the multiple subjects in this strike are fighting for.

The writers' union WGA demanded during negotiations that the AI not be allowed to receive bylines and that writers not be required to make changes based on what the AI has written, as this would also significantly reduce the number of hours worked. At the same time, producers are not allowed to train AI on scripts by union members without permission. Negotiations between the directors' union and the actors' union have also focused on this point, with the former asking for confirmation that AI cannot replace directors' union members in the performance of their duties, and the latter stating that the use of actors' likenesses for AI training without permission is prohibited.

Hollywood's attitude toward AI has been ambiguous. Some actors have allowed AI to clone their voices after they die, such as James Earl Jones, the voice of Darth Vader in the "Star Wars" series. But put into the broader film and television industry, the impact of technology needs to be clearly seen in a broader perspective to avoid infringing on the interests of different groups.

It's hard to predict when this strike will end; the WGA's strike in 1988 lasted 153 days, and the 2007 strike, which lasted from November to February of the following year, before the two sides reached an agreement. It is worth mentioning that both strikes were also due to the emergence of new technologies, the first being video tapes, the second DVDs, and this time AI.

As Brian Arthur concludes in The Nature of Technology, the economy responds to the emergence of new bodies of technology by changing the way activities are conducted, the composition of industries, and institutional arrangements, that is, the economy changes its structure in response to new bodies of technology. The new technologies will stir up the old patterns and naturally bring new problems and conflicts, which will be an eternal topic. No matter how it ends, the outcome of this negotiation will also change the industry forever, just like the previous two.

Big language models are a typical example. Before Sam Altman and his Open AI did not become famous, not only domestic, the whole AI circle also just as a new type of tool, investors began to become cautious, so much so that some of them ran to look at the new consumer, the reason is simple, compared to the complex technology, consumer goods category to better understand some.

We all know the story later. The marriage between Open AI and Microsoft forcibly renewed the fire of the originally cold AI industry, and the heat spread to the country, with unprecedented discussions inside and outside the circle.

Especially in the first half of this year, following the release of Baidu Wenxin Yiyin, all kinds of new and old companies have emerged, in addition to hearing every now and then that a certain company has released its own big model, the most delightful, there are new group chats born every day, among which there are those selling tutorials, those peddling private councils under the name of AI, and of course, there are also some serious talk about the current situation and future of big models.

The most paradoxical thing is that the four little dragons of AI, which had a certain fame in the AI circle in the past, are actually not the top stream this time.

The company has also launched other big models, unlike Open AI, which continues to strengthen the LLM, and is in a hurry to prove itself to the outside world.

But things are not as they should be, since 2022, Shang Tang has been repeatedly reduced by major shareholders, including Softbank Group and Alibaba, the former has reduced its holdings 4 times, with cash in excess of HK$326 million, and the latter has reduced its holdings 3 times. Some industry insiders believe that the reduction of major shareholders' holdings, for Shang Tang, whose revenue is currently declining and not yet profitable, is tantamount to sending an unpromising signal to the market.

This is also the point that this article wants to talk clearly: whether the big model windfall can bring new life to Shang Tang?

01 The dilemma of artificial intelligence, the big model can not change

On the other hand, Ali itself is also adjusting its business lines, so it needs to cut the side investment projects.

But these reasons still seem a bit far-fetched, because from the point of view of value investment, the reduction of holdings can only explain a problem: the probability is not a high-quality asset in the eyes of others.

Let's take Shang Tang Technology as an example, a company that has been holding the golden key since the day it started. If you put a label on it, there is no doubt that it is a scientist himself. Shang Tang's co-founder, Xiaogu Tang, is a professor at the Chinese University of Hong Kong and is considered by outsiders to be the pioneer and pathfinder of global face recognition technology.

According to the new eye incomplete statistics, the establishment of four years of Shang Tang financing has been more than 1.7 billion U.S. dollars, is then the world's largest financing, the highest valuation of artificial intelligence unicorn company. But the good times have not lasted long, since Shang Tang went public at the end of 2021, its market value has sunk all the way down, and now Shang Tang's market value hovers around HK$70 billion.

So here's the question, why did Shang Tang, once the rage, change its flavor after the IPO?

On this issue, there is a high praise answer on Zhihu: "The biggest problem of AI four small dragons is not the lack of clear business, but the initial development route is not clear, resulting in incoherent business, the previous technology failed to form effective precipitation, can not well help the newly proposed main business strategy. In other words, the AI four dragons are only now finding a clear development direction, and the sunk costs have not been transformed into nutrients, most of which have been wasted."

This comment was published two years ago, but even now, it is still not outdated.

Translated into words we can all understand, it means that AI is cold and capital is no longer obsessed with AI myth. Since the previous billions of dollars and nearly ten rounds of investment have not been able to make it hold to a larger market, now even if you have a new strategy and release a new product, in the eyes of the market, it is still necessary to put some discount.

From another perspective, this is actually the main reason why artificial intelligence has been lukewarm over the years.

The company's business perspective alone, if a technology does not find the right landing scene, then the technology is likely to be only a lonely appreciation. For a long time in the past, a number of star technology companies, including Shang Tang, over-obsessed with technical beliefs, ignoring the scene landing, performance in the business, that is, their tentacles are very long, whether it is the C-side, or B-side and G-side, as long as there is a suitable job, it will do.

After all, between the ideal and reality, in order to continue to tell the story of AI, these traditional software companies can do the work, they will probably continue to do.

02 Business soup problem is a typical AI industry problem

On the matter of AI gold content, there are three main judgment factors in the industry: R&D investment, revenue scale and growth rate, and net profit margin of main business, but people tend to pay attention to the first two and ignore the most critical last one.

Because about the first two, almost every company doing software can do not bad, the key is that many outsourcing, integrators neck point, is unable to form a scale effect and technical barriers, here is a technical misunderstanding, many people mistakenly believe that the technical barriers represent the number of patents owned by the enterprise, but the most practical measure should be, this technology can be applied to social scenarios, whether the lack of it is not.

In this regard, Open AI is a typical example. To this day, what the industry cares more about than the big models that have been launched, is actually how it actually builds the models and how it conducts model training.

This is also the most lacking competitiveness of some domestic AI companies.

In other words, the problem of Shangtang is not only a problem of Shangtang itself, but also an industry-wide problem.

This also explains why artificial intelligence is difficult to form an absolute barrier. According to IT Orange data, as of 2020, 30% of growing AI companies have not yet been invested, and many of these uninvested companies have not found a segmented value segment, and the competitive advantage of product differentiation is not obvious, and there is even serious homogeneous competition.

So the question arises, can the big model actually solve the current AI dilemma?

The answer is no, in essence, most of the big models now can not be called the real AGI. some industry insiders told the New Eye, "the measure of success of the big model is not just how many parameters, but what kind of scenario problem it can solve, and this scenario problem, with AI to solve lower cost, higher security. "

According to this logic, the current big model is still far from being able to support a main business, but in turn has exacerbated the black box of AI, originally everyone for artificial intelligence is already very puzzling, and now steeply launched a variety of new programs, its reliability and commercial value is even more questionable.

AI status quo is also roughly the same, in recent years bursts of fire and silence in the middle of the table is typical. As far as domestic players are concerned, you will find that the ones who are living well are basically stuck in vertical areas, such as KDDI in the field of AI voice, FanSoft in the field of intelligent BI, etc., but currently there is no run out of a giant similar to Microsoft or Snowflake-type.

03 Seemingly but not the wind mouth, is making the situation more and more confusing

After the big model exploded, many people thought it would be a super windfall.

But the fact is that Open AI next door has worked closely with Microsoft to try to integrate AI capabilities into the original Microsoft product system, and its Azure cloud computing business, office 365, and even its search business, Bing, have all undergone major upgrades.

However, the domestic AI environment is different.

Basically the giants, including Ali and Tencent, are more willing to develop their own big models rather than cooperate with other vendors, which is determined by the domestic Internet development path. China is a super market, both typical AI companies, or Internet companies with some R&D capabilities, they are more willing to close their doors to do it themselves, as for the ecology, most still remain in the sales perspective and verbal.

In this case, often exacerbate the degree of industry involution, so much so that there is a strange phenomenon, the big model more and more fire, while the positioning of artificial intelligence companies in turn more and more ambiguous. This is also another issue that needs to be thought about, according to Keynesian economic logic, the key to the domestic AI cold is that the supply far exceeds the real demand, and want to cultivate this market, still need time to settle and multiple efforts.

According to this logic, we really should let AI cool down, back to the rational track.

"After arriving at the company, we first answer emails, then check industry information and financing information, especially the AIGC track where new concepts appear every day, in addition to reading the latest academic papers from international academic journals. The biggest workload of this is checking industry research reports (research reports for short)." Xiaohui Wang, Managing Partner of Shengjing Jiacheng Investment, described his daily morning work rhythm to Burning Jiayuan.

Previously, if she wanted to obtain data on the scale of a certain industry, Wang needed to go through at least 5-6 research reports that were often dozens of pages long. In order to provide data to support the project, investors often need to spend more time looking for available data in the huge database of research reports.

Now "train AI to work for you", AIGC provides investors with efficiency tools.

"Now it's all AI+, I will use ChatGPT to highlight research reports and train ChatGPT to work for me." Wang told Burning Times that from the end of 2022 to the present, using AI tools to grab industry research reports and generate mind maps has multiplied her team's work efficiency. In addition, using ChatGPT to output basic code and take over design requirements has also saved Wang Xiaohui's efforts and costs.

Like Wang Xiaohui, Xiao, who is currently working in an overseas dollar fund, also said that in the overseas workplace, the AI tool is also a "magic tool for writing emails" to correct grammatical problems and expressions, and even some foreign investment institutions are trying to "let ChatGPT play the role of Soros, Warren Buffett, etc. to assist investment decisions. Some foreign investment institutions are even trying to "let ChatGPT play the role of Soros, Warren Buffett, etc. to assist investment decisions".

In addition to the auxiliary work, the AIGC fever has also brought about a change in the focus of investors.

At present, major domestic manufacturers such as Baidu, Tencent Youtu, Alibaba, Racer, Byte Jump, Netease, Shang Tang, Meitu, etc. have entered the AIGC track, including Baidu's Wenxin Yiyin, Tencent's writing robot "Dream Writer", Alibaba's AI online design platform Lubanner, etc., followed by AI leaders such as Wang Huiwen, Wang Xiaochuan, Li Zhifei, etc. also brought up the AI big model startup fever, AIGC, has become a hot area that people have to pay attention to investment.

Investors flocked to the AIGC investment track also appeared "good projects do not seek to invest" situation. A senior investor told Burning Sub-Year, "some investor friends and I complained, Wang Xiaochuan do big model, he took a large sum of money to invest were rejected."

At the same time, investors are looking for financing opportunities for good projects, but on the other hand, they need to be wary of being "cut leeks" by the AIGC entrepreneurs who rubbed the heat. But after researching, they found that they just copied the open source software of ChatGPT."

"Now many startups, back-end access to ChatGPT, front-end made a UI design, on the shelves of the Apple Store to boast that they are doing AIGC startups." Unknown Capital Managing Director William Wong said that this "batch of skin" AIGC project, "no technical barriers and business logic, just rubbing the heat.

Google CEO Eric Schmidt once said, "If someone offers you a place on a rocket, then, never mind the location is good or bad, go up first."

From web3 in 2022 to AIGC in 2023, investors are scrambling to get on the AIGC "express" train, after all, no one wants to fall in line under the boom.

But at present, there are still many challenges and problems to be solved in the domestic AIGC track, and even investors who are eager to get on the AIGC train are very cautious. Enterprise data shows that by the end of May, only 5 financing cases were completed in the field of AI text/image/video generation in 2023.

Under the wind mouth, investors are also inclined to rational, "now the concept of domestic AIGC is still on the early side, the wind mouth receded in the market whether there are still entrepreneurs continue to persist, is what we want to see." William added.

Crowding into the AI track, mud and sand

With the boom, investors are turning their attention to the AI track.

"Investors and entrepreneurs all over Silicon Valley are moving in this direction now." Jiang Tao, founder of CSDN and founding partner of Geek Gang Ventures, said bluntly at the Artificial Intelligence Conference Pioneer Session (GAIDC).

Under the boom, AIGC investment track also presents a hot, good projects have to be "grabbed", as well as the mud and sand, "rubbing the hot" projects and many other very different phenomena.

"Dabbling in the heat of more, too few good projects, encounter good projects, investors often rely on 'grab'." Wang Xiaohui told Burning Sub-Yuan.

As early as before 2018, Wang has been concerned about the AIGC field, spare early layout, she can also feel, with the end of 2022 ChatGPT triggered a buzz, driving the AI track investment projects completely "robbed", betting on the final good projects that can run out, become the goal of investors.

Wang added, "In any industry, there are not many good projects in the first place, and good projects are not worried about financing."

Even, as early as 2018, when Wang Xiaohui participated in a research institute exchange activity, he paid attention to the research shared by Zhang Linfeng, who was still a doctoral student at that time, about "AI for science", "his research was very peak technological innovation at that time and even now. That's why I expressed my investment intention early on, but he hadn't returned to China yet."

Wang Xiaohui said, wait until the epidemic during Zhang Linfeng returned to China to found the deep potential technology, a good project also triggered other large capital competition, and finally after several rounds of fighting, Wang Xiaohui only in the angel round of investment in deep potential technology, since then, deep potential technology and access to high tide, warp and woof, Qiming, Huawei Hubble and other institutions of multiple rounds of investment.

In this regard, Oasis Capital founding partner Zhang Jinjian said that with the influx of startups into the AIGC track, as of March 2023, the past three weeks, Oasis's team can only sleep four or five hours a day, a day to see close to 20 projects.

When we searched "activity line" and other platforms, ChatGPT and AIGC were on the top of the hot search list, and the AIGC summits, large and small, were also attended by major investment institutions.

Xiao said that in the financial sector, there will always be new hot spots, and with ChatGPT on fire, "companies will also ask us to look at the AI track," and just in the Japanese market she is responsible for, she also observed that "many AI companies that were originally listed and half-dead, after a fire many big funding came in."

Like Xiao, William told Burning Times that although he had previously been focusing on the AI track, he had mainly concentrated on face recognition and natural language processing directions, and the AIGC fever brought up by ChatGPT made him start to adjust his focus, "Now I will look more at the direction of chatbots and new application scenarios based on AI big models. "

Whether it is the early layout of the AI track investors, or just turn their attention to the later, busy in the "chaotic" AI track gold, but also because some of the hot startups "headache".

"Now many startups, in fact, no technical innovation at all, just access to ChatGPT, front-end packaging a little, do a UI design, on the shelves of the Apple Store to sell." william bluntly said.

This "batch of skin" AIGC project, the lack of technical barriers and mature business logic, more just to rub the heat for profit, "this project often only need to ask a few words, such as what is your algorithm? Where is the technical barrier? And then you can try to find out." william said.

"To digital people, for example, now the domestic start-up projects are saying they are original development, in fact, many (projects) are not at all, and end up spelling low price, application scenarios are also limited, there is not much to see." A small red book blogger @昊昊来过, who previously worked for a domestic dual-currency fund, told Burning Next Gen, "There is no lack of funding for the AI track, just that many good projects are now overseas."

And perhaps a more direct way to distinguish the startups that are rubbing it in is to distinguish the founders.

Wang Xiaohui told Burning Times, "In the past, capital liked entrepreneurs from big factories, but now capital prefers serial entrepreneurs, after all, past experience does not lie, and in the field of hard technology investment, we pay more attention to whether the founders have the ability to make disruptive technological innovations and industrialize these technological innovations. In this way we pay attention to their scientific background, whether they have published relevant papers in top international scientific journals, whether they are the lead author of the papers, whether the team has the ability to commercialize, etc."

"But in the AIGC field, we focus more on end-to-end models and want more people who are familiar with the industry to be able to translate AI capabilities into efficient tools." Wang added.

Training AI to work for itself

Under ChatGPT set off a big AI model startup boom, the value of AIGC, which burns hot, to assist investors in their work comes more directly and easily.

"For me, ChatGPT is simply the most efficient search portal for finding research reports (research reports) at the moment." William wong, Managing Director of Unknown Capital, told Burning Subgenre that as an investor, a necessary task every day is to find industry research reports, but in the past, to do in-depth research on an industry, one needed to switch multiple browsers, as well as brokerage websites, to download more than 10 research reports.

"After searching with ChatGPT, it can automatically grab the research reports, which can save me at least one to two days' time." william wong said excitedly.

"Investors all have to learn to train ChatGPT to work for themselves." Wang told Burning Times, "With ChatGPT you can take out industry research reports that are often dozens of pages and hundreds of pages, grab out the key points and form a mind map."

Wang graduated in 1991 from one of the few economics and management majors that had access to investment knowledge at that time. In the era without office software, Wang even needed to spend 2-3 weeks to build financial models, but now with AI tools, it also shortens the long document preparation time in the past.

And in addition to using ChatGPT to work for themselves, investors are also using AI to replace parts of their work such as design and code writing.

"ChatGPT is able to replace part of the design manpower." Wang said that because one of the data sources used to train ChatGPT is Github, which is the largest open source community for software, some common programs only need to input the needed functions, and ChatGPT can output the basic code, and many software companies are already using ChatGPT to improve the efficiency of software developers.

"Using AI to replace repetitive mechanical work can also help us reduce some of our expenses." Wang said.

Like Wang and William, Xiao said that in addition to using AI for document summarization, ChatGPT has become her "personal secretary" in the overseas English communication environment, "ChatGPT can help proofread grammatical problems and correct expressions. It is very useful for writing emails."

"In fact, ChatGPT can also be used to play the role of Soros, Warren Buffett, etc. It is very logical and can assist in decision making." @HaoHao came by to add.

Recently, according to the results of an experiment released by finder.com, a British financial advisory website, the net value of stock portfolios constructed from 38 listed companies recommended by ChatGPT rose about 4.9% in five weeks from March 6, outperforming 10 popular fund products recommended by Interactive Investor, a British online investment platform (average return Some European and American quantitative investment institutions are also trying to use ChatGPT to build investment strategies.

While we don't know the real numbers, we have heard from reliable sources that the supply of computing power is so tight that demand is more than 10 times greater! So we believe that access to computing resources at the lowest total cost of ownership is now a determining factor in the success of AI companies.

In fact, we have seen many companies spend more than 80% of their total funding on compute resources.

In this article, we try to break down the cost factors for AI companies. The absolute numbers will of course change over time, but we otherwise I AI companies are limited by access to computing resources that will immediately ease. So, hopefully, this is a helpful framework to think about.

Why are the computational costs of AI models so high?

There is a wide variety of generative AI models, and the cost of inference and training depends on the size and type of model. Fortunately, most of today's most popular models are based on Transformer architectures, including popular Large Language Models (LLMs) such as GPT-3, GPT-J, or BERT. While the exact number of inference and learning operations of a transformer is model-specific (see this paper), there is a fairly accurate rule of thumb that depends only on the parameter The number of parameters (i.e., the weights of the neural network) of the model and the number of input and output Tokens.

Token is basically a short sequence of a few characters. They correspond to words or parts of words. The best way to get an intuition about tokens is to try tokenization using publicly available online tokenizers such as OpenAI. For GPT-3, the average length of a token is 4 characters.

Transformer's rule of thumb is that for a model with an input of p parameters and an output sequence of length n tokens, forward pass-through (i.e., inference) requires approximately 2np floating-point operations (FLOPS)¹. For training the same model, each token requires approximately 6*p floating-point operations (i.e., the additional backward pass requires four more operations ²). You can estimate the total training cost by multiplying this by the number of tokens in the training data.

The memory requirements of the Transformer also depend on the model size. For inference, we need p model parameters to fit in memory. For learning (i.e., backpropagation), we need to store additional intermediate values for each parameter between the forward and backward passes. Assuming we use 32-bit floating point numbers, this is an additional 8 bytes per parameter. For training a model with 175 billion parameters, we need to keep more than a terabyte of data in memory -- more than any GPU in existence today, requiring us to partition the model onto different cards. The memory requirements for inference and training can be optimized by using shorter length floating point values, with 16 bits becoming commonplace and 8 bits expected in the near future.

The table above shows the size and computational cost of several popular models. GPT-3 has about 175 billion parameters, corresponding to 1,024 token inputs and outputs, and a computational cost of about 350 trillion floating-point operations (i.e., Teraflops or TFLOPS). Training a model like GPT-3 requires about 3.14*10^23 floating point operations. Other models such as Meta's LLaMA have much higher computational requirements. Training such a model is one of the more computationally demanding tasks humans have undertaken to date.

To summarize: AI infrastructure is expensive because the underlying algorithmic problems are extremely difficult to compute. The algorithmic complexity of sorting a database table with a million entries is trivial compared to the complexity of generating a single word with GPT-3. This means that you have to choose the smallest model that solves your use case.

The good news is that for transformer, we can easily estimate how much computation and memory a model of a particular size will consume. Therefore, choosing the right hardware becomes the next consideration.

The time and cost debate for GPUs

How does computational complexity translate into time? A processor core can typically execute 1-2 instructions per cycle, and due to the end of Dennard Scaling, processor clock rates have remained stable at around 3 GHz for the last 15 years. Executing a single GPT-3 inference operation without utilizing any parallel architecture would require 350 TFLOPS/(3 GHz*1 FLOP) or 116,000 seconds, or 32 hours. This is highly impractical; instead, we need specialized chips to accelerate this task.

Virtually all of today's AI models run on cards that use a large number of dedicated cores. For example, the NVIDIA A100 graphics processor has 512 "tensor cores" that can perform 4×4 matrix multiplication (equivalent to 64 multiplications and additions, or 128 FLOPS) in a single cycle. Artificial intelligence gas pedal cards are often referred to as GPUs (graphics processing units) because the architecture was originally developed for desktop gaming. In the future, we expect AI to increasingly become a distinct product family.

With a nominal performance of 312 TFLOPS, the A100 could theoretically reduce GPT-3 inference time to about 1 second. However, this is an overly simplified calculation for a number of reasons. First, for most use cases, the bottleneck is not the computational power of the GPU, but the ability to get data from dedicated graphics memory to the tensor core. Second, 175 billion weights would take up 700 GB and would not fit into the graphics memory of any GPU. Techniques such as partitioning and weight streaming would need to be used. Third, there are some optimizations (e.g., using shorter floating-point representations such as FP16, FP8, or sparse matrices) that are being used to speed up the computation. Overall, however, the figures above give us an intuitive idea of the overall computational cost of LLM today.

It takes about three times as long to train a transformer model per token as it does to perform inference. However, given that the training dataset is 300 million times larger than the inference cues, training takes a billion times longer. On a single GPU, training takes decades; in practice, this is done on large compute clusters in dedicated data centers or, more likely, in the cloud. Training is also harder to parallelize than inference because updated weights must be swapped between nodes. memory and bandwidth between GPUs often becomes a more important factor, and high-speed interconnects and dedicated architectures are common. For training very large models, creating a suitable network setup may be the primary challenge. Looking ahead, AI gas pedals will have networking capabilities on the card or even on the chip.

So, how does this computational complexity translate into cost? As we saw above, a GPT-3 inference, which takes about 1 second on the A100, has a raw computational cost of between $0.0002 and $0.0014 for 1000 tokens (compared to OpenAI's pricing of $0.002/1000 token). This is a very low price point, making most text-based AI use cases economically viable.

Training GPT-3, on the other hand, is much more expensive. At the above rate, again calculating only the computational cost of 3.14*10^23 FLOPS, we can estimate the cost of a single training session on the A100 card to be $560,000. In practice, for training, we will not get nearly 100% efficiency on the GPU; but we can also use optimization to reduce training time. Other estimates of GPT-3 training costs range from $500,000 to $4.6 million, depending on hardware assumptions. Note that this is the cost of a single run, not the overall cost. Multiple runs may be required, and cloud providers will want a long-term commitment (more on this below). Training top-tier models is still expensive, but affordable for well-funded startups.

In summary, today's generative AI requires significant investment in AI infrastructure. There is no reason to believe this will change in the near future. Training a model like GPT-3 is one of the most computationally intensive tasks ever undertaken by humans. While GPUs are becoming faster and we are finding ways to optimize training, the rapid expansion of AI offsets both of these effects.

AI Infrastructure Considerations

At this point, we have tried to give you some idea of the scale required to perform AI model training and inference, and the underlying parameters that drive them. With this background, we would now like to provide some practical guidelines on how to decide which AI infrastructure to use.

External vs. internal infrastructure GPUs are cool. Many engineers and engineering-minded founders favor configuring their own AI hardware, not only because it allows for fine-grained control over model training, but also because there is some fun to be had in leveraging large amounts of computational power (Appendix A).

However, the reality is that many startups -- especially app companies -- do not need to build their own AI infrastructure on day one. Instead, hosted modeling services like OpenAI or Hugging Face (for language) and Replicate (for image generation) allow founders to quickly search for product-market fit without having to manage the underlying infrastructure or models.

These services have become so good that many companies can depend on them directly. Developers can achieve meaningful control over model performance through cue engineering and higher-order fine-tuning abstractions (i.e., fine-tuning through API calls). Pricing for these services is consumption-based, so it is also often cheaper than running separate infrastructure. We have seen a number of application companies generating over $50 million in ARR, valued at over $1 billion, running hosted modeling services in the background.

On the other hand, some startups -- especially those training new base models or building vertically integrated AI applications -- can't avoid running their own models directly on GPUs. Either because the models are actually products and the team is looking for "model-market fit," or because fine-grained control over training and/or inference is required to achieve certain functionality or reduce marginal costs at scale. Either way, managing the infrastructure can be a source of competitive advantage.

Building the Cloud and Data Center In most cases, the cloud is the right place for your AI infrastructure. For most startups and large companies, the lower upfront costs, ability to scale up and down, regional availability, and fewer distractions from building your own data center are attractive.

There are, however, a few exceptions to this rule:

If you have a very large operation, running your own data center may become more cost effective. The exact price point varies depending on location and setup, but typically requires more than $50 million per year in infrastructure spending. You need very specific hardware that you can't get from a cloud provider. For example, there are no widely available types of GPUs, and unusual memory, storage, or networking requirements. You can't find an acceptable cloud for geopolitical reasons. If you do want to set up your own data center, there are already comprehensive GPU price/performance analyses available for your own setup (e.g., Tim Dettmer's analysis). In addition to the cost and performance of the cards themselves, the choice of hardware depends on power, space, and cooling. For example, two RTX 3080 Ti cards together have similar raw compute power to the A100, but each consume 700 W vs. 300 W. Over a three-year lifecycle, at a market price of $0.10/kWh, the 3500 kWh power difference adds nearly two times the cost of the RTX3080 Ti (~$1,000).

In summary, we expect the vast majority of startups to use cloud computing.

Compare Cloud Service Providers Amazon Web Services (AWS), Microsoft Azure and Google Cloud Platform (GCP) all offer GPU instances, but new providers have emerged that focus specifically on AI workloads. Here is the framework we have seen used by many founders to select a cloud provider:

Pricing: The table below shows pricing for some of the major and smaller specialty clouds as of April 7, 2023. This data is for reference only, as instances vary greatly in terms of network bandwidth, data egress costs, additional costs for CPUs and networks, available discounts, and other factors.

The computing power of a particular piece of hardware is a commodity. To put it bluntly, we would expect prices to be fairly uniform, but that's not the case. While there are substantial feature differences between clouds, they are not enough to explain the nearly 4x difference in pricing between vendors for the on-demand NVIDIA A100.

At the top of the price range, large public clouds charge a premium based on brand reputation, proven reliability, and the need to manage a variety of workloads. Smaller, specialized AI providers offer lower prices by running dedicated data centers, such as Coreweave, or by nesting other clouds, such as Lambda Labs.

In practice, most large buyers negotiate prices directly with cloud providers, often committing to some minimum spend requirement as well as a minimum time commitment (we are seeing 1-3 years). After negotiations, the price differences between clouds narrow somewhat, but we see the rankings in the table above remain relatively stable. It is also important to note that smaller companies can get aggressive pricing from a professional cloud without a large spending commitment.

Availability: The most powerful GPUs, such as the Nvidia A100 s, have been in short supply for the past 12+ months.

Considering the huge buying power and resource pools of the top three cloud providers, it is logical to assume they have the best availability. But, somewhat surprisingly, many startups don't find this to be true. The big cloud providers have a lot of hardware, but also a lot of customer demand to meet -- Azure, for example, is the primary host for ChatGPT -- and are constantly adding/releasing capacity to meet that demand. Meanwhile, Nvidia has committed to making hardware widely available across the industry, including allocating it to new specialty providers. (They are doing this both to be fair and to reduce their dependence on a few large customers who are also competing with them.)

As a result, many startups are finding more available chips at smaller cloud providers, including the cutting-edge Nvidia H100 s. If you are willing to work with newer infrastructure companies, you may be able to reduce the wait time for hardware and potentially save money in the process.

Compute delivery models: Today's large clouds only offer instances with dedicated GPUs because GPU virtualization is still an unresolved issue. Specialized AI clouds offer other models, such as containers or batch jobs, that can handle individual tasks without incurring the startup and teardown costs of instances. If you are comfortable with this model, it can significantly reduce costs.

Network interconnection: Specifically for training, network bandwidth is a major factor in choosing a provider. Training certain large models requires clusters that use dedicated networks between nodes, such as NVLink. for image generation, exit traffic costs are also a major cost driver.

Customer Support: Large cloud providers serve a large number of customers in thousands of product SKUs. Unless you are a large customer, it can be difficult to get customer support attention or get issues resolved. Many specialized AI clouds, on the other hand, offer fast and responsive support for even small customers. This is partly because they operate on a smaller scale, but also because their workloads are more homogeneous, so they have more incentive to focus on specific AI features and bugs.

Comparing GPUs All else being equal, the highest-end GPUs will perform best on almost all workloads. However, as you can see in the table below, the best hardware is also quite expensive. Choosing the right type of GPU for your particular application can significantly reduce costs and may make the difference between a viable and non-viable business model.

Deciding how far to go down the list - i.e., determining the most cost-effective GPU choice for your application - is primarily a technical decision that is beyond the scope of this article. However, we will share below some of the selection criteria that we believe are most important:

Training and Inference: As we saw in the first section above, training a Transformer model requires us to store 8 bytes of data for training, in addition to model weights. This means that a typical high-end consumer GPU with 12 GB of memory can barely be used to train a model with 4 billion parameters. In practice, training large models is done on clusters of machines, ideally with many GPUs per server, lots of VRAM, and high bandwidth connections between servers (i.e., clusters built with top-of-the-line data center GPUs).

Specifically, many models are most cost effective on the NVIDIA H100, but for now it is hard to find and usually requires a long-term commitment of more than a year. The NVIDIA A100, on the other hand, runs most model training; it is easier to find, but for large clusters, it may also require a long-term commitment.

Memory requirements: The number of parameters for large LLMs is too large to fit on any card. They need to be partitioned into multiple cards and require a training-like setup. In other words, even for LLM inference, you may need H100 or A100. but smaller models (e.g., Stable Diffusion) require less VRAM. while A100 is still popular, we have seen startups use A10, A40, A4000, A5000, and A6000, and even RTX cards.

Hardware support: While the vast majority of workloads at the companies we talked to are running on Nvidia, some companies are starting to experiment with other vendors. The most common is Google's TPU, and Intel's Gaudi 2 seems to be getting some attention. The challenge with these vendors is that the performance of your model is often highly dependent on the availability of software optimizations for these chips. You may have to do a PoC to get an idea of performance.

Latency requirements: In general, less latency-sensitive workloads (e.g., batch data processing or applications that do not require interactive user interface responses) can use less powerful GPUs. this can reduce compute costs by a factor of 3-4 (e.g., compare A100 s vs. A10 s on AWS). User-facing applications, on the other hand, often require high-end cards to deliver an attractive real-time user experience. Optimizing the model is often necessary to bring the cost to a manageable range.

Peak: Generative AI companies often see a sharp rise in demand because the technology is so new and exciting. It is not uncommon for requests to increase by a factor of 10 in a day, or consistently by 50% per week, on top of a new product release. It is often easier to handle these spikes on low-end GPUs because more compute nodes are likely to be available on demand. If this traffic comes from less engaged or less retained users, it often makes sense to serve such traffic with lower cost resources at the expense of performance.

Optimization and scheduling models Software optimization can dramatically impact the runtime of a model - a 10x gain is not uncommon. However, you need to determine which methods will work best for your particular model and system.

Some techniques work for a fairly wide range of models. Speedups achieved using shorter floating point representations (i.e. FP16 or FP8 compared to the original FP32) or quantization (INT8, INT4, INT2) are usually linear in the reduction of the number of bits. This sometimes requires model modification, but there are now a growing number of techniques to automate work with mixed or shorter precision. Pruned neural networks reduce the number of weights by ignoring low values of weights. Combined with efficient sparse matrix multiplication, this can achieve significant speedups on modern GPUs. In addition, another set of optimization techniques addresses memory bandwidth bottlenecks (e.g., by streaming model weights).

Other optimizations are highly model-specific. For example, Stable Diffusion has made significant progress in the amount of VRAM required for inference. Still another class of optimizations is hardware-specific. NVIDIA's TensorML includes some optimizations, but can only be run on NVIDIA hardware. Last, but not least, the scheduling of AI tasks can create huge performance bottlenecks or improvements. Assigning models to GPUs to minimize weight swapping, picking the best GPU for the task if multiple GPUs are available, and minimizing downtime by batching workloads ahead of time are common techniques.

Finally, model optimization remains a black magic, and most of the startups we've spoken to have partnered with third parties to help solve some of these software aspects. Typically, these are not traditional MLops vendors, but companies that specialize in optimizing for specific generative models, such as OctoML or SegMind.

In this article, we will discuss the role of heroes in movies and explore some of the most iconic heroic characters in film history. We will also examine the appeal of heroism in movies and the strategies filmmakers use to create compelling heroes. Finally, we will discuss how you can get involved in our community and receive a free NFT by subscribing to our account.

The Role of Heroes in Movies

Heroes play a critical role in movies, serving as the central figures around which the story revolves. They often embody certain values or ideals, such as justice, honor, or sacrifice, and serve as role models for audiences. Heroes can take many forms, from the swashbuckling adventurers of classic Hollywood to the complex anti-heroes of modern cinema. Some of the most iconic heroic characters in film history include:

1. Indiana Jones:

Indiana Jones is the beloved adventurer from the Indiana Jones franchise. He is known for his wit, charm, and bravery, as well as his trademark fedora and leather jacket.

2. Superman:

Superman is the iconic superhero from DC Comics, known for his incredible strength, speed, and ability to fly. He is often portrayed as a symbol of hope and justice.

3. Luke Skywalker:

Luke Skywalker is the hero of the original Star Wars trilogy, known for his journey from humble farm boy to Jedi Knight. He embodies the values of courage, perseverance, and selflessness.

The Appeal of Heroism in Movies

Heroism in movies has a powerful appeal, as it taps into our deepest hopes and desires. Heroes embody qualities that we aspire to, such as courage, selflessness, and determination. They inspire us to be our best selves and to believe that anything is possible. Some of the most significant reasons why heroism in movies is so appealing include:

1. Escapism:

Movies offer a way to escape from the stresses and challenges of everyday life. Heroes take us on epic journeys and adventures, allowing us to forget our problems and immerse ourselves in a thrilling world of action and adventure.

2. Inspiration:

Heroes inspire us to be our best selves and to believe that we can overcome any obstacle. They embody qualities that we admire, such as courage, selflessness, and determination, and serve as role models for audiences.

3. Catharsis:

Watching heroes overcome challenges and triumph over evil can be a cathartic experience, allowing us to release pent-up emotions and feel a sense of satisfaction and closure.

Join Our Community: How to Get Involved

If you are interested in exploring the role of heroes in movies and sharing your insights and analysis with others, we invite you to join our community. Our voluntary group is dedicated to studying the power of heroism in film and sharing our knowledge and insights with others.

As a special offer, we are also giving away free NFTs to our subscribers. NFTs, or non-fungible tokens, are digital assets that represent ownership of a unique item, such as a piece of art or a collectible. By subscribing to our account, you can mint your own free NFT, which will give you exclusive access to our community and our insights into the world of heroes in movies. You'll also receive regular updates and analysis on the latest movies and TV shows featuring heroic characters.

https://opensea.io/assets/ethereum/0x5B4057A070c4b9d6376E7c7c421f843B4D029933/0

Whether you're a casual movie fan or a dedicated scholar of film, our community is a great place to explore the power of heroism in movies and share your passion with others. By joining us, you'll have the opportunity to connect with like-minded individuals and to gain a deeper understanding of the role of heroes in film.

In conclusion, heroes have been a fixture of movies since the earliest days of cinema. They embody the qualities that we admire most, such as courage, selflessness, and determination, and serve as role models for audiences. Whether we're watching a classic adventure film or a modern superhero blockbuster, heroes capture our imaginations and inspire us with their bravery and perseverance. If you're someone who loves studying the movies industry and wants to explore the power of heroism in film, we invite you to join our community and share your insights and analysis with us. By subscribing to our account, you can also mint your own free NFT and gain exclusive access to our community and insights.

The tech press has been in a frenzy lately over the arrival of ChatGPT-4. On Tuesday, March 14 (EST), OpenAI announced the official launch of GPT-4, a large multimodal model that it says can take image and text input and output text "more creatively and collaboratively than any previous version " and "solves puzzles more accurately because it has a broader range of common sense and problem-solving capabilities."

II. The innovation of ChatGPT-4

Before understanding the innovation of ChatGPT-4, it is important to start by understanding artificial intelligence, while many of us are still struggling with the question "What is artificial intelligence?" Many of us are still struggling with the question, "What is AI? The world of artificial intelligence has moved on in a big way, and now AI can do all sorts of things for us.

But what exactly is ChatGPT and why has it caused such a stir around the world? ChatGPT's own answer is this: ChatGPT is an artificial intelligence (AI) tool designed to communicate with humans in a natural conversational way. It understands human language and generates responses that look human. Think of it as having a conversation with a truly intelligent computer program: it can answer all your questions, provide information, and talk to you. It can be used in chatbots, virtual assistants and other applications that require human communication.

In short, the current version of ChatGPT is an AI chatbot that uses the GPT-4 language model. While it's not perfect, it can quickly generate human-like text responses based on any prompt you provide. Investors in Microsoft ChatGPT have even tried to use it in products like Word, Powerpoint and Bing, so the AI can help people write better documents, help people make better presentations and improve refinements based on contextual search results.

Introduced this week, GPT-4 is a large multimodal model that accepts image and text input and then outputs correct text responses. GPT-4 achieves leaps and bounds in the following areas: powerful image recognition; an increase in text input limit to 25,000 words; significant improvements in response accuracy; and the ability to generate lyrics, creative text, and implement stylistic changes.

Meanwhile OpenAI says it has partnered with several companies to incorporate GPT-4 into their products, including Duolingo, Stripe and Khan Academy. the GPT-4 model will also be available as an API for subscribers to the paid version of ChatGPT Plus. Developers can sign up and build apps with it. Microsoft has said that the new Bing (Bing) search engine will run on top of the GPT-4 system. chatGPT-4 king bomb upgrade to make artificial intelligence once again become the focus of attention.

Third, so ChatGPT-4 will affect how the Web3 industry?

Web3 is a rapidly growing ecosystem of decentralized applications built on blockchain technology that promises to revolutionize various industries. Integrating ChatGPT-4 into Web3 could have a significant impact on the development of decentralized applications and the overall user experience.

Combining the views of various authorities and individuals, here are our pragmatic views on how ChatGPT-4 may impact Web3.

Improving the user experience Integrating ChatGPT-4 into Web3 applications can dramatically improve the user experience. With its ability to understand natural language queries, ChatGPT-4 can make it easier for users to interact with decentralized applications, even if they are not familiar with the underlying technology.

For example, ChatGPT-4 can be used to create conversational wallets that allow users to execute transactions by simply stating their intent in natural language, reducing the barrier to entry for Web3. This helps make the Web3 ecosystem more accessible to a broad audience, further driving its mass adoption.

Chat-based Wallets Wallets are the primary entry point for interacting with DAPPs in Web3. Just as ChatGPT-4 was used as a fundamental construct to re-architect the user experience in Web2 applications, a similar trend can be postulated for crypto wallets. The new wallet experience based on ChatGPT-4 can be described as a chat-like approach using natural language where users can easily communicate their intent to request information, perform a transaction, or just perform a specific task.

Natural Language Based Block Browsers In Web3, browsers are the primary way to search and interact with blockchains, but unfortunately, the user interface of these browsers is geared towards blockchain experts rather than the average user. Using ChatGPT-4, the average user can ask questions like "Are there large institutions transferring funds to Binance?" , "When was the last time something like this happened?" or "Are there any interesting patterns in recent trading activity?" and other natural language descriptions.

The user's search experience can be significantly upgraded with technologies such as ChatGPT-4, for example, for browsers. The new generation of Microsoft Bing is powered by artificial intelligence.

Intelligent NFT Models such as ChatGPT-4 enable a new generation of NFT that includes conversational intelligence. imagine a version of your favorite NFT series that allows you to ask questions about a creator's inspiration or specific artistic details.

Smart Contract Development Assistant Writing smart contract programs is a complex and specialized task that requires a great deal of expertise on the part of the developer. However, with ChatGPT-4's Codex feature, Solidity code can be generated using natural language descriptions. Imagine having a smart contract assistant that can produce code snippets for a specific task, such as "requesting a flash loan in Aave," by simply typing the query in natural language.

Smart Contract Security Testing While smart contract security testing is critical, smart contract auditing is known to be a slow, expensive and tedious process. The audit process relies heavily on running tests that are often not well understood or significant to smart contract developers. Future web3 smart contract security testing could use a specialized version of ChatGPT-4 for smart contract auditing, capable of receiving linguistic input and executing a set of tests on a given smart contract.

IV. Summary

The integration of ChatGPT-4 into Web3 applications is expected to have a significant impact on the development and overall user experience of decentralized applications. With its ability to understand natural language queries, ChatGPT-4 can make the Web3 ecosystem more accessible to a broader audience. As more and more projects adopt ChatGPT and Web3, we are likely to see an explosion of innovation across industries. This will lead to exciting new products and services that will revolutionize industries such as gaming, finance, education, etc. The future is promising with the combination of two revolutionary forces, AI + web3.

Look at the new projects coming into Web3 - they are far more likely to choose to build on competing L1s or Rollups.

This article will explain why this is happening.

There was a time when Ether started to focus on being a "robust currency". A number of improvement proposals were used to achieve this, including EIP-1559, "mergers" and a focus on deflating ETH all played a role.

These proposals directly benefit ETH holders by adjusting the token economics of ETH to derive more value from network fees ......

while aiming to reduce the supply of ETH in the process - leading smart investors to hold and pledge only ETH.

The crypto community is largely dominated by traders/investors, so these initiatives are often welcomed. People want ETH to add value (even if they use Ponzi-economy tactics)

But the goal of the business is to maximize value for the stakeholders.

New builders of Ether who understand this principle are starting to realize that Ether has just created a toxic environment for their protocols.

Why the path of Ether (especially after the merger) will continue to drive founders to choose other blockchains.

Ether will eventually become a replacement for Bitcoin's store of value, and not a very good one at that.

Those who think a deflationary base token will benefit ecosystem activity must never have taken an economics class.

Here's why.

Smart contract layers try to be the foundation of the economy, using their tokens to fuel activity.

The economy needs inflation to promote growth.

If you don't believe me, read @CryptoHayes' recent article: https://entrepreneurshandbook.co/a-samurai-a-knight-and-a-yankee-211bf975a31d

Deflationary tokens only lead to people buying and "HODLing".

It discourages investment in the broader ecosystem and imposes an extremely high opportunity cost on anyone who tries to buck the trend ......

This is well documented in world economics.

For a more modern example, look at the impact of deflation on the Japanese economy.

Compare this to the Federal Reserve's (designed to promote growth) target inflation rate.

In the course of human history, deflationary models have never been a boost to the economy ......

So, why has Bitcoin survived?

Because bitcoin is a store of value. It's not something that builds an economy, it's just a digital commodity ...... just like gold.

However, for an economy, you need currency. The US economy relies on dollars, not gold.

This is why the deflationary model is so detrimental to ethereum.

Ether was created as a smart contract layer for the digital economy.

Now, investors need to ask themselves why invest in growth (e.g. new protocols) when they can hold and pledge ETH ......

It's getting worse ......

Due to the adjusted token economics of Ether - especially in a post-merger environment - all activity on the blockchain accumulates back into base tokens and the supply of base tokens decreases.

Using the previous example, imagine if you tried to force an economy on top of gold (with a reduced supply), but anything you did in that economy would make gold more valuable.

Why would anyone choose to spend their gold?

This is the token model of Ether going forward.

There is an additional, fatal step ......

If the protocol founders leave Ether faster, the opportunity cost of just holding and pledging ETH for smart people to get invested is too high ......

So where does the cost of driving ETH appreciation and reducing supply come from?

Yes ...... so that sucks!

The problem is, I'm not the only one who knows this sucks.

Here are the main alternatives to this dilemma, which is why these alternatives are a lifesaver

1. "Rollup will solve this problem"

No, they won't. People who say this address many broken assumptions.

They start with the assumption that Rollups/Layer-2 on ETH2 will somehow reduce the cost of execution with deflationary base tokens

This is never mathematically possible. In this reality, everyone will eventually be priced out. Try again.

Second, they assume that all Rollups are willing to pay rent to Ether for the security model of ETH.

This is just silly.

In business, if you have two providers offering you similar products at different costs, you will choose the cheaper one every time.

Also ...... remember that businesses have a responsibility to maximize value for their stakeholders?

In this case, why didn't Rollup finally decide to create its own chain?

Oh wait, that already happened.

Importantly, Rollups can lead to severe liquidity dispersion - one of the most serious problems with cryptocurrencies.

The more Rollups there are, the worse it gets and breaks the user experience.

Cross-chain bridge solutions do not address this Rollup problem (especially with slower EVM chains). There is an unavoidable composability/security tradeoff.

OK, so start from scratch with the Rollup idea ......

There is a way for the protocol founders to survive deflationary base tokens on ETH L1 via treasury management!

Let's look at this crazy hypothesis.

1. Hedge Funds

Here ...... protocol founders should keep their funds in ETH so that their income can keep pace with ETH and provide investors with a more appropriate risk/reward profile ......

Do I need to explain the downside of this?

Therefore, in order to move beyond Ether, founders must.

a) engage in active money management, which basically means becoming a hedge fund and taking advantage of market timing ......

or

b) go they have to go to another chain to capture growth.

Yes ...... go to another chain. Separate the value from Ether. Push the founders out of Ether.

That's what we said at the beginning.

It features superstar Steph Curry going through a goofy version of his era - eating cereal, making pasta, carving ice sculptures - while narrator Shaquille O' Neal) insists that Curry knows everything there is to know about encryption. An irritated Curry repeatedly denies it.

"I'm not an expert, and I don't need to be," Curry finally says to the camera, holding up the FTX app on his phone. "

With FTX, I have everything I need to safely buy, sell and trade cryptocurrencies.

Give Curry and FTX some honesty.

This new ad says what should be obvious to anyone paying attention: the countless celebrities who have joined the crypto (and NFT) trend almost certainly know little about the products they sell.

This goes beyond the inherently transactional nature of corporate sponsorship. Everyone knows that athletes advertise because they are paid to do so, not because they actually use the product.

On the other hand, if one knows that Curry - or Tom Brady, Paris Hilton, Charli D'Amelio, Snoop Dogg or Matt Damon) - could explain what someone is buying when they invest in cryptocurrencies, it would be shocking.

The honesty of the Curry ad is offset by its cynicism, which sets a new standard for an industry with a lot of idle.

The crypto-ad blitz began late last year, rising in tandem with the price of digital assets. The Super Bowl infamously showcased several of the industry's big-budget ads.

The most notable was an FTX ad featuring comedian Larry David, who saw cryptocurrency as a passing fad with a "don't be like Larry" kicker.

As some observers have pointed out, these ads conspicuously ignore any substantive merits of crypto. Instead, they try to instill a sense of FOMO, or fear of missing out, by suggesting that viewers who don't buy now will regret it as much as Larry does.

These FOMO ads at least leave open the possibility that consumers will learn about cryptocurrencies before investing. The curry ads eschew this pretense.

To be fair, there is a difference between an expert who doesn't do something and one who knows nothing about it. But the ad is clearly aimed at people who are hesitant to invest in cryptocurrencies because they don't know anything about them.

The message to them is: don't worry, Stephen doesn't worry either! And perhaps, by extension, neither does anyone else! If everyone else is operating in ignorance, maybe you're not at any great disadvantage. So go ahead, trade away. fTX did not respond to a request for comment.

Warren Buffett, the legendary investor, is said to have advised, "Never invest in a business you can't understand.

Traditional investing is a bet that the business you invest in will become more valuable over time. As the fundamentals improve, the business grows and becomes more profitable, more people will be willing to pay more for a portion of it, thus increasing the value of your stock.

If you can't understand how a business makes money, you have no reason to make a reasonable judgment about the performance of its stock. However, as the Curry ad shows, there are ways to avoid the intermediate steps in the crypto market.

Forget the fundamentals: For the average investor, the decision to buy a certain cryptocurrency seems to be a pure bet that others will want to spend more money on it in the future. This is the spirit behind the modal stock phenomenon, which is philosophically closer to the crypto world than to traditional stock market investing.

For investments, there is a name whose value depends entirely on finding future buyers willing to pay more than you put in: Ponzi schemes.

Critics have used this label to label crypto for almost as long as crypto has existed.

More recently, the criticism has been bolstered by an unlikely source. the founder and CEO of FTX, Sam Bankman-Fried, appeared on the Odd Lots podcast last week.

During the discussion, Bloomberg Finance columnist Matt Levine asked Bankman-Fried to explain "income farming," a form of crypto investment in which people can buy "liquidity pools" that can pay ultra-high interest rates, but can also be rushed south.

As an example, Bankman-Fried asks Levine to imagine "a company that makes boxes" that, despite some noble marketing rhetoric, does nothing.

Bankman-Fried went on to say that the company could issue tokens that people would buy - although "so far, we haven't given a compelling reason why the box would make any money. Finally, the company could promise to distribute more tokens to anyone who puts money in the box, that is, anyone who lends them money. bankman-Fried says that in this case, the buzz around the token could give it a market cap of $20 million. (Although, as Levin points out, it should be zero. Then, as more and more people see that the box is paying out valuable tokens, they'll want to put more money into it, thus increasing the value even further.

"It's like, it's a valuable box, as evidenced by all the money people apparently decided should be in the box," Bankman-Fried said. "Who are we to say they're wrong?"

"I consider myself a rather cynical person," Levin replied. "It's a lot more cynical than the way I would describe agriculture. You just think, 'Well, I'm in a Ponzi scheme, and that's fine.

Surprisingly, Bankman-Fried didn't try very hard to fight back. "I think it's a pretty reasonable response," he says. "I think there's a frustrating validity to it.

What's more frustrating is what happened next in the crypto market: nothing.

You might think that the head of a major crypto trading platform speaking out loud about the quiet part, admitting on a popular podcast that he is at least partially in the Ponzi business, might lead to some sort of crisis or loss of confidence in the industry.

That's not the case. It shows that people don't really care if they put their money in a box that does nothing, as long as they think they'll be one of the groups shelling out more money later.

In this sense, it may be wrong to compare the crypto market to the stock market. Gambling is probably a more appropriate comparison. A slot machine doesn't do anything either; it's effectively a box that takes money in and spits it out.

As with gambling, you can avoid risk by not putting your money in the bottom pool. But actually opting out of the crypto economy is becoming increasingly difficult.

Even if you don't invest yourself, you may be living in a crypto world. Retirement funds don't bet on the Super Bowl, but two public pension funds in Fairfax County, Virginia, invested in crypto funds last fall and are now reportedly considering getting into income agriculture.

Wall Street banks have long been skeptical of cryptocurrencies, reluctantly entangling themselves more deeply in them, such is their fear of missing out on a big payday.

It's not just the economy that is increasingly wrapped up in the fate of crypto.

The industry's rise has spawned a generation of overnight millionaires and billionaires, some of whom have political ambitions.

As The Washington Post recently reported, crypto investors and executives are pouring millions of dollars into the upcoming midterm elections, trying to help elect candidates who support the industry's preferred regulations.

Then there's Bankman-Fried, who is worth an estimated $24 billion at age 30 and has been increasing his DC footprint. His political donations seem to have more to do with pandemic readiness than crypto-friendly regulation; a practitioner of effective altruism, Bankman-Fried has vowed to give almost all of his money to causes that will have the greatest impact.

So that's one thing crypto boxes do: transfer money from someone else's bank account to his so he can direct it as he sees fit.

He seems to be betting that he can do more good with other people's money than those people can. The rest of us will have to hope that he and the other crypto mobsters are right. After all, we are not experts.