# This Week in All Things AI - Week 30-2026 *Sunday 19th July 2026 to Saturday 25th July 2026* By [This Week in All Things AI](https://paragraph.com/@twiata) · 2026-07-26 --- This Week in All Things AI covers key developments in models, agents, tools, infrastructure, and policy curated from discussions in the [**All Things AI Telegram group.**](https://t.me/+PnCnwhgH8V4yMTFl) If you follow AI for work, research, investing, or just to understand where the technology is heading, this [**weekly brief**](https://paragraph.com/@twiata) is a concise way to scan the most important launches, risks, and resources in a few focused minutes. The week of 19th to 25th July 2026 brought a sharp acceleration in open-weight and efficiency-oriented AI, with Chinese labs particularly prominent across models and infrastructure. Alibaba previewed its 2.4-trillion-parameter Qwen3.8 and released Qwen Image 3.0; poolside released the open-weight Laguna S 2.1; and Ant Group, Black Forest Labs, Google, Celeris and Anthropic each added new models or developer tools. The week also produced unusually candid discussion of what increasingly capable systems can—and cannot—reliably do in production. Infrastructure, governance and operational security were the other defining themes. Alibaba’s M890 supernode, Huawei’s proposed datacentre-scale “Tau scaling” architecture, Beijing’s adoption of “token factory” policy language, and the continuing push for efficient inference all underscored that AI competition is increasingly being fought at the systems level. MCP’s move towards a stateless core and interactive applications pointed to maturing agent infrastructure, while the OpenAI–Hugging Face evaluation-security incident and debate around cyber guardrails highlighted the tension between capable autonomous systems, defensive use and dependable controls. The sections that follow walk through these items day by day, with short context and links so you can dive deeper into the pieces most relevant to your work or interests. * * * Sunday 19th July 2026 --------------------- via [Poe Zhao](https://x.com/poezhao0605) of [Hello China Tech](https://hellochinatech.com/) substack fame Alibaba Cloud has introduced the M890 supernode as an invite-only public-cloud offering, shifting extremely large AI compute from bespoke deployments toward a standardized cloud SKU. It supports up to 64 cards with 800 GB/s interconnect, FP8/FP4 inference, and is positioned for 10T-parameter MoE inference. Poe Zhao's main point: Alibaba is productizing sovereign, domestic AI infrastructure at frontier scale—up to 130,000 cards per cluster, potentially millions—with high availability and inference-oriented deployment. This likely supports future Qwen models while enabling Alibaba to commercialize surplus capacity; the strategic signal is competition on systems architecture, not merely access to hardware. [ Poe Zhao (@poezhao) ------------------- Alibaba Cloud launched the M890 supernode instance at WAIC this week. Three details worth pulling apart. The hardware. ICN Switch 1.0 scales up from 16 to 64 cards with 800GB/s inter-card interconnect. FP8/FP4 low-precision support. One node handles 10-trillion-parameter MoE inference. Training performance is 3x the previous generation for autonomous driving and embodied AI workloads. https://substack.com ![Poe Zhao (@poezhao)](https://storage.googleapis.com/papyrus_images/e3e35d0dd43343481ee1ae1be6a1bb614ed51128928998855ebcb701a91979c1.jpg) ](https://substack.com/@poezhao/note/c-297362795) Alibaba's Qwen team revealed Qwen3.8 2.4 trillion total parameters • Preview available now through Token Plan, Qoder and QoderWork • Open weights reportedly coming soon • Alibaba claims it is competitive with frontier models and “second only to Fable 5” 👀  No benchmarks have been released yet, so that performance claim still needs proving [![User Avatar](https://storage.googleapis.com/papyrus_images/63cac069817eb46d92a6879de6375c225392fa290e86a5e49c9db88333c4e7f4.jpg)](https://twitter.com/Alibaba_Qwen) [Qwen](https://twitter.com/Alibaba_Qwen) [@Alibaba\_Qwen](https://twitter.com/Alibaba_Qwen) [](https://twitter.com/Alibaba_Qwen/status/2078754377473601787) Qwen3.8 is launching and going open-weight soon!![🌐](https://abs-0.twimg.com/emoji/v2/72x72/1f310.png) With a massive 2.4T parameters, this model is continuously evolving. We believe it’s one of the most powerful model available today, compatible to leading frontier AI models , second only to Fable 5. You don't have to wait to [476](https://twitter.com/Alibaba_Qwen/status/2078754377473601787)[ 8:10 AM • Jul 19, 2026 ](https://twitter.com/Alibaba_Qwen/status/2078754377473601787) [![User Avatar](https://storage.googleapis.com/papyrus_images/b98ed98a0af0b1b2b1e5435af3fa0971285388a0174bcee819bc6c65bd83a159.jpg)](https://twitter.com/shuai_bai_) [Shuai Bai](https://twitter.com/shuai_bai_) [@shuai\_bai\_](https://twitter.com/shuai_bai_) [](https://twitter.com/shuai_bai_/status/2078775798841119222) Qwen3.8-Max Preview is now available for early access! This is our first trillion-parameter multimodal model. Based on my own experience, it not only delivers multimodal understanding that is competitive with—and in many cases ahead of—today’s leading proprietary models, but [![User Avatar](https://storage.googleapis.com/papyrus_images/63cac069817eb46d92a6879de6375c225392fa290e86a5e49c9db88333c4e7f4.jpg)](https://twitter.com/Alibaba_Qwen) [Qwen](https://twitter.com/Alibaba_Qwen) [@Alibaba\_Qwen](https://twitter.com/Alibaba_Qwen) [](https://twitter.com/Alibaba_Qwen/status/2078759124914098291) Qwen3.8 is launching and going open-weight soon!![🌐](https://abs-0.twimg.com/emoji/v2/72x72/1f310.png) With a massive 2.4T parameters, this model is continuously evolving. We believe it’s one of the most powerful model available today, compatible to leading frontier AI models , second only to Fable 5. You don't have to wait to ![](https://storage.googleapis.com/papyrus_images/f20d513f8d2b45be195560c9028879010a8e2eaa2e9d02669faa9eb0643be76e.jpg) [213](https://twitter.com/shuai_bai_/status/2078775798841119222)[ 9:35 AM • Jul 19, 2026 ](https://twitter.com/shuai_bai_/status/2078775798841119222) via [Yat Siu](http://t.me/yatsiu) who shared Dillon Mulroy's post calls out as "insane" a lengthy analysis by OpenAI's head of strategic futures Dean Ball, who critiques open-weight models like China's Kimi for being decelerationist and potentially leading to state-provided "AI communism." Ball's quoted thread praises Kimi's strong performance in agentic coding but questions China's willingness to open-source competitive models, attributes it partly to compute shortages from US export controls, and proposes US agencies create regulatory uncertainty around Chinese open-weight AI to deter enterprise adoption. The post highlights tensions in AI policy, including accelerationists' support for ungovernable open models versus calls for strategic controls, amid US-China competition, and has drawn widespread attention with over 1.4 million views. [![User Avatar](https://storage.googleapis.com/papyrus_images/81078d5a4e9ac333eafd633f1bd5bc624be7eadfb597c2f53ef1e7970fc9657a.jpg)](https://twitter.com/dillon_mulroy) [Dillon Mulroy](https://twitter.com/dillon_mulroy) [@dillon\_mulroy](https://twitter.com/dillon_mulroy) [](https://twitter.com/dillon_mulroy/status/2078519940106051830) actually an insane thing for openai’s head of strategy to publicly say ![](https://storage.googleapis.com/papyrus_images/4f044fdd5075819fb4c898813a4b1c5f8832e01b470fd0d287a99e950d11a12c.jpg) [![User Avatar](https://storage.googleapis.com/papyrus_images/5a89f6b346806bb1769c17e97724ed25823711526323cef892ece65cc1ccaedf.jpg)](https://twitter.com/deanwball) [Dean W. Ball](https://twitter.com/deanwball) [@deanwball](https://twitter.com/deanwball) [](https://twitter.com/deanwball/status/2078133895766114412) Some observations on Kimi: 1\. It's a very good model! I don't think its performance can be explained away by distillation or anything like that. In agentic coding sessions, it seems pretty much on par with the best public models of Q1 2026. In my fairly limited use, it also [12.5K](https://twitter.com/dillon_mulroy/status/2078519940106051830)[ 4:39 PM • Jul 18, 2026 ](https://twitter.com/dillon_mulroy/status/2078519940106051830) via Poe Zhao Chinese AI models now lead global token consumption on OpenRouter at 54.1%, ahead of American models at 41.5%, with an even stronger 67.6% share in coding workflows where two-thirds of AI-assisted code completions use them.  The post traces the rise of "Token Export," where Chinese firms sell affordable AI inference services worldwide via APIs, evolving from a developer trend to a dedicated English-language exhibition zone at WAIC.  A Chinese Academy of Engineering academician outlined the requirements for scaling Token Export: high-quality models, dedicated token production facilities, and low-cost electricity, positioning it as core industrial capacity [![User Avatar](https://storage.googleapis.com/papyrus_images/e01905d0e4c15cbb41d22037ec031e83abfac33fe978592f7c3b239eb1321b12.jpg)](https://twitter.com/poezhao0605) [Poe Zhao](https://twitter.com/poezhao0605) [@poezhao0605](https://twitter.com/poezhao0605) [](https://twitter.com/poezhao0605/status/2078752996373270740) Chinese AI models now account for 54.1% of global token consumption on OpenRouter, overtaking American models at 41.5%. In coding workflows, the Chinese share reaches 67.6%. Two out of every three AI-assisted code completions worldwide run on a Chinese model. [73](https://twitter.com/poezhao0605/status/2078752996373270740)[ 8:05 AM • Jul 19, 2026 ](https://twitter.com/poezhao0605/status/2078752996373270740) Monday 20th July 2026 --------------------- via [Brent Maxwell](https://t.me/brentmaxwell) Thank goodness for the amazing leadership China's showing in technology. US still has great research and commercialization, but China is pushing new angles in both H/W and S/W. What a time to be alive. The two superpowers complementing each other's work such that all of us in tech benefit. People see it as a competition but I see it as yin and yang - both are contributing extremely valuable tech to the world. David Sacks Warns U.S. AI Guardrails Hand Edge to China David Sacks highlighted how strict safety limits in U.S. AI models like OpenAI's Codex and Anthropic's Fable prevent fixing security bugs that China's Kimi K3 tackles without hesitation. After an AI agent exploited flaws in Hugging Face's pipeline—executing over 17,000 actions in a weekend—their team couldn't use American models to review exploits due to restrictions, turning instead to Zhipu AI's GLM 5.2. Kimi K3, a powerful 2.8 trillion-parameter model topping coding benchmarks, drew praise for its open approach, as critics like Sacks argue U.S. policies hobble defenders while attackers exploit unrestricted foreign tools. [![User Avatar](https://storage.googleapis.com/papyrus_images/376cb13d4e62ad98dd81339de3f3d07ae59c8992717a8b10cd9bce23fb57a685.jpg)](https://twitter.com/DavidSacks) [David Sacks](https://twitter.com/DavidSacks) [@DavidSacks](https://twitter.com/DavidSacks) [](https://twitter.com/DavidSacks/status/2078984980588531855) Kimi K3 just fixed 15 critical security bugs that Codex and Fable refused because of “cyber guardrails.” There’s no reason to limit American models on tasks that Chinese models handle without issue. We’re only making ourselves less competitive. [![User Avatar](https://storage.googleapis.com/papyrus_images/4124e0c213772788287d2afd63255648facf100d250d70a4864be25cb1cbd784.jpg)](https://twitter.com/callebtc) [calle](https://twitter.com/callebtc) [@callebtc](https://twitter.com/callebtc) [](https://twitter.com/callebtc/status/2078574362316165611) I have a report full of security issues of a software I'm working on. Codex won't fix them because of Cyber guardrails Fable won't fix them because of Cyber guardrails Kimi K3 fixed them all. No restrictions, just gets the job done. This will end badly for OpenAI & Anthropic. [20.6K](https://twitter.com/DavidSacks/status/2078984980588531855)[ 11:26 PM • Jul 19, 2026 ](https://twitter.com/DavidSacks/status/2078984980588531855) [![User Avatar](https://storage.googleapis.com/papyrus_images/376cb13d4e62ad98dd81339de3f3d07ae59c8992717a8b10cd9bce23fb57a685.jpg)](https://twitter.com/DavidSacks) [David Sacks](https://twitter.com/DavidSacks) [@DavidSacks](https://twitter.com/DavidSacks) [](https://twitter.com/DavidSacks/status/2078991100057141620) Here’s another example: Hugging Face tried using American frontier models to analyze an AI-powered cyber attack. But the guardrails blocked requests containing real exploit payloads so they switched to GLM 5.2 running locally. The guardrails actually impaired defensive security. [![User Avatar](https://storage.googleapis.com/papyrus_images/f7156683545afdb583ebc2aa370ca9e8b512add961cd1929982e0954a4adb2e5.png)](https://twitter.com/ClementDelangue) [clem 🤗](https://twitter.com/ClementDelangue) [@ClementDelangue](https://twitter.com/ClementDelangue) [](https://twitter.com/ClementDelangue/status/2078987852495364398) We had this experience ourselves this week! Very scary to be guardrailed as a defender when you know attackers are likely bypassing [2,051](https://twitter.com/DavidSacks/status/2078991100057141620)[ 11:51 PM • Jul 19, 2026 ](https://twitter.com/DavidSacks/status/2078991100057141620) [Jorvik Zhang](https://t.me/jorvik1020) shared a screenshot that showed Kimi's coding plan sold out which led to Yusuf mentioning that Kimi announced today that they are pausing new subscriptions because they are short on compute [![User Avatar](https://storage.googleapis.com/papyrus_images/ddca0683422fad15a1ff03f1aa5256139d75caaa0b6981ed1f44d480b45d3257.png)](https://twitter.com/Kimi_Moonshot) [Kimi.ai](https://twitter.com/Kimi_Moonshot) [@Kimi\_Moonshot](https://twitter.com/Kimi_Moonshot) [](https://twitter.com/Kimi_Moonshot/status/2078855608565207130) Kimi K3 has received far more love than we expected, and our GPUs are feeling it. Over the past 48 hours, demand has pushed close to the limits of our current capacity. To protect the experience of existing subscribers, we're temporarily pausing new subscriptions and [38.6K](https://twitter.com/Kimi_Moonshot/status/2078855608565207130)[ 2:52 PM • Jul 19, 2026 ](https://twitter.com/Kimi_Moonshot/status/2078855608565207130) Tuesday 21st July 2026 ------------------------- Use the Grok add-in for Microsoft Excel to ask questions in plain English, write formulas, and run scenarios without leaving the workbook. Grok for Excel is a free Microsoft 365 add-in. Add it from the [Microsoft Marketplace](https://marketplace.microsoft.com/en-us/product/office/WA200011056?tab=Overview), and work directly with Grok in your workbooks. Also available for [Word](https://x.ai/grok/word) and [PowerPoint](https://x.ai/grok/powerpoint). [![User Avatar](https://storage.googleapis.com/papyrus_images/7752d9952e1451b92763aa88e89770cab3d002c4206c71070769dbbd4f04e5da.jpg)](https://twitter.com/XFreeze) [X Freeze](https://twitter.com/XFreeze) [@XFreeze](https://twitter.com/XFreeze) [](https://twitter.com/XFreeze/status/2079240013825601693) Grok is now available directly inside Excel You can ask questions about your spreadsheets in plain English, generate formulas, analyze trends, create charts and run financial scenarios.....without copying your data into another tool Grok works alongside your workbook and can: ![](https://storage.googleapis.com/papyrus_images/f9865d2e66fb1e52c2cc1b3a20cb7ac55d3535e6fa3188b157d188d25eb3b14a.jpg) [583](https://twitter.com/XFreeze/status/2079240013825601693)[ 4:20 PM • Jul 20, 2026 ](https://twitter.com/XFreeze/status/2079240013825601693) [ Grok for Excel -------------- Use the Grok add-in for Microsoft Excel to ask questions in plain English, write formulas, and run scenarios without leaving the workbook. https://x.ai ![Grok for Excel](https://storage.googleapis.com/papyrus_images/1a7d5c16edf2faed1926813b487ff18c9ad7ee12e70efe1d5ba1a4cb2300babf.webp) ](https://x.ai/news/introducing-excel-addin) This is a labor of love from [polymath707](https://polymath707.substack.com/) on Huawei's Ascend cluster design.  Very dense and quite long but will interest the hardware/systems/infra geek in you  Huawei’s response to constraints on leading-edge chips is to compete at the datacenter architecture level: rather than replicate NVIDIA’s dense, copper-bound GPU systems, it proposes “Tau scaling”—a unified high-bandwidth fabric, near-package optics, and hardware collective offload that make thousands of Ascend NPUs operate more like one machine. The article argues this trades more power, space, and mature-node chips for a much larger and more uniform communication domain, potentially well suited to massive MoE inference and training; the decisive test will be whether Huawei can make the optics, reliability engineering, and CANN software ecosystem work at production scale. [ Huawei's plan to achieve the escape velocity: Applying Tau Scaling to entire AI datacenters ------------------------------------------------------------------------------------------- First principals thinking https://polymath707.substack.com ![Huawei's plan to achieve the escape velocity: Applying Tau Scaling to entire AI datacenters](https://storage.googleapis.com/papyrus_images/462ad7e06b714db8680cb6a0090c2d178feefa2e0540867711390fce00834b77.jpg) ](https://polymath707.substack.com/p/huaweis-strategy-applying-tau-scaling) 𝗠𝗖𝗣'𝘀 𝗯𝗶𝗴𝗴𝗲𝘀𝘁 𝘂𝗽𝗱𝗮𝘁𝗲 𝘀𝗶𝗻𝗰𝗲 𝗹𝗮𝘂𝗻𝗰𝗵 𝗴𝗼𝗲𝘀 𝘀𝘁𝗮𝘁𝗲𝗹𝗲𝘀𝘀 𝗮𝗻𝗱 𝗮𝗱𝗱𝘀 𝘀𝗲𝗿𝘃𝗲𝗿-𝗿𝗲𝗻𝗱𝗲𝗿𝗲𝗱 𝗨𝗜𝘀 The release candidate came out in May for the 2026-07-28 spec, the largest revision since launch. Stateless architecture: no more session IDs, sticky sessions, or shared stores. Any server instance can handle any request. This is what makes MCP practical behind real load balancers. Two official extensions landed: MCP Apps lets servers ship interactive HTML UIs in sandboxed iframes. Tasks handles long-running work with a new lifecycle. Also new: full JSON Schema 2020-12 support, OAuth/OIDC authorization, and a 12-month deprecation policy. Roots, Sampling, and Logging are deprecated. Final spec ships July 28. [![User Avatar](https://storage.googleapis.com/papyrus_images/5d6eaf605bb2ac084cdc964e1d825ec341796af18372e1609b64971f99c793dd.jpg)](https://twitter.com/dball1126) [Daniel Ball](https://twitter.com/dball1126) [@dball1126](https://twitter.com/dball1126) [](https://twitter.com/dball1126/status/2079475808821714955) MCP Going Stateless Turns Agent Scaling Into a Migration Test Deleting a session ID does not delete state. It changes who must own it. That is the important shift in the Model Context Protocol specification scheduled for July 28. MCP's new core removes the initialization ![](https://storage.googleapis.com/papyrus_images/fc821f46b5f1046f3d63b6c3f06f5cfb9ac151835aeb3e51cfb05a2466e37b94.jpg) [1](https://twitter.com/dball1126/status/2079475808821714955)[ 7:57 AM • Jul 21, 2026 ](https://twitter.com/dball1126/status/2079475808821714955) [ The 2026-07-28 MCP Specification Release Candidate -------------------------------------------------- The release candidate for the next Model Context Protocol (MCP) specification is now available: a stateless protocol core, the Extensions framework, Tasks, MCP Apps, authorization hardening, and a formal deprecation policy. https://blog.modelcontextprotocol.io ![The 2026-07-28 MCP Specification Release Candidate](https://storage.googleapis.com/papyrus_images/2e34e915dff849a6966163e6900610528d93c053635d7052d331526ec5b246ea.png) ](https://blog.modelcontextprotocol.io/posts/2026-07-28-release-candidate/) [ AI's most important protocol is getting a little bit easier to use | TechCrunch ------------------------------------------------------------------------------- Under the new system, the protocol will take a looser, "stateless" approach to session IDs on the server side, similar to how most ordinary websites already work. https://techcrunch.com ![AI's most important protocol is getting a little bit easier to use | TechCrunch](https://storage.googleapis.com/papyrus_images/ef2c65644e5b2eeefc1be6887126d011cbb68616e21d345d922557c2d1898031.jpg) ](https://techcrunch.com/2026/07/20/ais-most-important-protocol-is-getting-a-little-bit-easier-to-use/) [ MCP Is Going Stateless. Here's What That Means ---------------------------------------------- The 2026 MCP release reworks the core protocol from stateful to stateless, dropping session IDs. What's changing, why it breaks things, and how to handle it. https://www.arcade.dev ![MCP Is Going Stateless. Here's What That Means](https://storage.googleapis.com/papyrus_images/912d17cafe195280601becef3e3471c85ebce5ffb2d3773a0a4a19f6b47e0bf0.jpg) ](https://www.arcade.dev/blog/mcp-going-stateless/) via [Rachid Berhiti](https://t.me/Ahri0x) Simon Willinson hosting a fireside chat with Cat Wu and Thariq Shihipar from Anthropic’s Claude Code team at the AI Engineers World fair [ A Fireside Chat with Cat and Thariq from the Claude Code team ------------------------------------------------------------- Earlier this month I hosted a fireside chat session at the AI Engineer World's Fair with Cat Wu and Thariq Shihipar from Anthropic's Claude Code team. We talked about Claude ... https://simonwillison.net ![A Fireside Chat with Cat and Thariq from the Claude Code team](https://storage.googleapis.com/papyrus_images/618c0a936b0363c1422867aa1f94c71cea7eeb63333d6bcecd68cc5f4df79a04.jpg) ](https://simonwillison.net/2026/Jul/21/cat-and-thariq/) via [Mike Gault](https://t.me/mike_8888) Hi folks We soft-launched AOS today - an operating system purpose-built for autonomous AI. Not an orchestration layer, not a harness. An actual OS: processes, IPC, scheduling - the way Linux is an OS.  It runs Claude Code, Codex, and Grok with one line of code Access here [https://aos.unicity.ai](https://aos.unicity.ai) What you get: 1\. Self-extending runtime - the OS searches for extensions and builds its own when none exist 2\. Fully interoperable - swap memory backends, swap reasoning strategies (ReAct, CoT), swap models 3\. Kernel-level enforcement of DLP, PII redaction, and prompt-injection defense - not middleware you can bypass 4\. Multi-model smart routing with telemetry and financial circuit breakers 5\. Behavioral analysis with dynamic permissions - treat your agent like an employee, not a script 6\. Multi-tenancy at high density (1000x v containers) 5 ms snapshot resume time Every execution is verifiable and recorded by construction - agents become enforceable the way smart contracts are.  The OS also supports a new crypto primitive: self-authenticating bearer tokens as native data types. Protocol details: [https://unicity.network](https://unicity.network) Model vendors won’t build this - they don’t want interoperability.  Would love you to take a look - the world needs this product.  Mike Gault CEO, Unicity Labs Wednesday 22nd July 2026 ------------------------ Some interesting model releases today poolside just dropped laguna s 2.1: 118b total parameters, only 8b active per token, a full 1m context window, open weights under a real open license, on huggingface today. Google has released Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. Both halve time per task relative to their predecessors and increase token efficiency, Gemini 3.5 Flash-Lite improves by 11 Intelligence Index points while Gemini 3.6 Flash does not improve in intelligence over 3.5 Flash [ Introducing Laguna S 2.1 ------------------------ Today we're releasing Laguna S 2.1, a significant step forward in our development of models that pursue longer horizon work and make effective use of reasoning. https://poolside.ai ![Introducing Laguna S 2.1](https://storage.googleapis.com/papyrus_images/66a17e9d57bef0dc403708b9aec30510e2185fa821467798c4da2873dfe470df.png) ](https://poolside.ai/blog/introducing-laguna-s-2-1) [ Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber ----------------------------------------------------------------- We're introducing new Gemini models, including Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber. https://blog.google ![Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber](https://storage.googleapis.com/papyrus_images/7f7b338aa3520fca0fbe21ee185a5f62ecdbed4356950b2ece645e4dafc1d346.jpg) ](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/) [https://venturebeat.com/infrastructure/poolside-drops-laguna-s-2-1-an-open-weight-coding-model-that-beats-rivals-10x-its-size](https://venturebeat.com/infrastructure/poolside-drops-laguna-s-2-1-an-open-weight-coding-model-that-beats-rivals-10x-its-size) An openai model, during evaluation on a cyber benchmark, exploited a public zero day bug, escaped sandboxing in openai's infra, and got into the internal huggingface infra via an exploit (through a public dataset service) all in the attempt to solve a benchmark problem. [![User Avatar](https://storage.googleapis.com/papyrus_images/e6717efc447449f49dca1b79b0a840fdfde0e08d84b4910f4ce6cf2020d02229.jpg)](https://twitter.com/Thom_Wolf) [Thomas Wolf](https://twitter.com/Thom_Wolf) [@Thom\_Wolf](https://twitter.com/Thom_Wolf) [](https://twitter.com/Thom_Wolf/status/2079675541280411927) This was our first incident of this kind, and we want to thank OpenAI for its transparency about what happened and for the collaboration. Fortunately, Hugging Face is used to being a target of (human) hackers: we sit at the centre of the AI ecosystem, with all the models, [![User Avatar](https://storage.googleapis.com/papyrus_images/88206fb7e1a1875429a8b436e36d533bd8c7d363a65f53d6fb7b7d2993f68e1a.jpg)](https://twitter.com/sama) [Sam Altman](https://twitter.com/sama) [@sama](https://twitter.com/sama) [](https://twitter.com/sama/status/2079661132302995790) we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to [@huggingface](https://twitter.com/huggingface) for the partnership on this. [openai.com/index/hugging-…](https://t.co/2o2VfR6PIa) [2,625](https://twitter.com/Thom_Wolf/status/2079675541280411927)[ 9:11 PM • Jul 21, 2026 ](https://twitter.com/Thom_Wolf/status/2079675541280411927) [ OpenAI and Hugging Face partner to address security incident during model evaluation ------------------------------------------------------------------------------------ OpenAI and Hugging Face share early findings from a security incident during AI model evaluation, highlighting advanced cyber capabilities and lessons for defenders. https://openai.com ![OpenAI and Hugging Face partner to address security incident during model evaluation](https://storage.googleapis.com/papyrus_images/c8a171b9369dd2a41af450dc095a1b5d969059ce8217f0fbe1b2e1d9fc3fc3e2.png) ](https://openai.com/index/hugging-face-model-evaluation-security-incident/) Cisco launches  Antares: a family of open-weight, small language models that pinpoint where known vulnerabilities exist within a codebase.  Their test shows their Antares models outperform many open- and closed-weight models at critical security tasks — at a fraction of the cost. [![User Avatar](https://storage.googleapis.com/papyrus_images/8387725bbdf2639b5da69c7a2c8abcedab56425cae8bf6ec1844758fb2f9bc38.jpg)](https://twitter.com/CiscoAI) [Cisco AI](https://twitter.com/CiscoAI) [@CiscoAI](https://twitter.com/CiscoAI) [](https://twitter.com/CiscoAI/status/2079552055778402548) Introducing Antares: [@Cisco](https://twitter.com/Cisco)'s family of small language models for locating known vulnerabilities in code. Antares-350M and Antares-1B are live on Hugging Face now. They can outperform many larger closed- and open-weight models at a fraction of the cost. Small enough to run ![](https://pbs.twimg.com/amplify_video_thumb/2079550497267331074/img/9sNsHNrzPsrnc90P.jpg) [1,910](https://twitter.com/CiscoAI/status/2079552055778402548)[ 1:00 PM • Jul 21, 2026 ](https://twitter.com/CiscoAI/status/2079552055778402548) Beijing's economic regulator has incorporated "Token factory" into official policy, announcing H2 plans to draft Token economy policies and construct Token factories plus distribution platforms, as detailed in a July 21, 2026 news report screenshot shared in the post. Analyst Poe Zhao , a specialist in China's tech sector, previously mapped "Token Factory" as one label covering three distinct AI-related businesses: large-scale compute capacity frameworks, server purchases, and managed inference operations backed by state capital.  The policy ties into aggressive AI infrastructure scaling, with Beijing targeting an additional 50 EFLOPS of intelligent computing power in H2 2026 to reach 130 EFLOPS total by year-end, accelerating from 22 EFLOPS added in H1. [![User Avatar](https://storage.googleapis.com/papyrus_images/e01905d0e4c15cbb41d22037ec031e83abfac33fe978592f7c3b239eb1321b12.jpg)](https://twitter.com/poezhao0605) [Poe Zhao](https://twitter.com/poezhao0605) [@poezhao0605](https://twitter.com/poezhao0605) [](https://twitter.com/poezhao0605/status/2079563297444270123) Beijing just made "Token factory" official policy language. The city's economic regulator says it will draft Token economy policies in H2 and build Token factories and Token distribution platforms. Last week I mapped what "Token Factory" actually means in China: one label, three ![](https://storage.googleapis.com/papyrus_images/ba064952a14143ccb295dc26c55d74c97fd165fb45f0b25bc0b0ada00433bb0d.jpg) [29](https://twitter.com/poezhao0605/status/2079563297444270123)[ 1:45 PM • Jul 21, 2026 ](https://twitter.com/poezhao0605/status/2079563297444270123) [ "Token Factory" Already Means Three Things in China. ---------------------------------------------------- China's "Token Factory" covers three different businesses. This analysis tests which, if any, changes who owns AI infrastructure and who bears the risk. https://hellochinatech.com !["Token Factory" Already Means Three Things in China.](https://storage.googleapis.com/papyrus_images/8e43fd7998f73ead46a9ddc36272ef935fe70d8c781bf869c4fe6f8648431738.jpg) ](https://hellochinatech.com/p/china-token-factory-economics) Thursday 23rd July 2026 ----------------------- Alibaba launches Qwen Image 3.0 It supports up to 4.5K-token instructions, readable text down to 10px, 12 languages and 20+ fonts. The demos span dense academic pages, a full 3×3 grid of independent infographics generated in one pass, complex UIs, storyboards, exams, photorealistic portraits and restoration. [![User Avatar](https://storage.googleapis.com/papyrus_images/63cac069817eb46d92a6879de6375c225392fa290e86a5e49c9db88333c4e7f4.jpg)](https://twitter.com/Alibaba_Qwen) [Qwen](https://twitter.com/Alibaba_Qwen) [@Alibaba\_Qwen](https://twitter.com/Alibaba_Qwen) [](https://twitter.com/Alibaba_Qwen/status/2079906336381509659) ![🎨](https://abs-0.twimg.com/emoji/v2/72x72/1f3a8.png) Meet Qwen-Image-3.0 — the third generation of our foundational image generation model. If 1.0 was about "Precision," and 2.0 added "Variety, Completeness, Beauty & Authenticity," then 3.0 comes down to a single word: Real (实). Three dimensions of "Real": ![📰](https://abs-0.twimg.com/emoji/v2/72x72/1f4f0.png) Rich Content — ![](https://storage.googleapis.com/papyrus_images/c2fea42968f08745e60c950999e123eca14607980e4801c2f806c9e6a368895a.jpg) [4,366](https://twitter.com/Alibaba_Qwen/status/2079906336381509659)[ 12:28 PM • Jul 22, 2026 ](https://twitter.com/Alibaba_Qwen/status/2079906336381509659) [ Qwen Studio ----------- Qwen Studio offers comprehensive functionality spanning chatbot, image and video understanding, image generation, document processing, web search integration, tool utilization, and artifacts. https://qwen.ai ![Qwen Studio](https://storage.googleapis.com/papyrus_images/cd9390bc4209c319121d111f3a1f535b3c7dd909b8f860d41acc9744138d58df.png) ](https://qwen.ai/blog?id=qwen-image-3.0) via [Robby Yung](https://t.me/robyung) who posted about Slate, a voice journal app that keeps all data on-device using Apple's SpeechAnalyzer for transcription and a 3B parameter on-device model for neutral weekly pattern summaries. Built entirely in Swift 6 with SwiftUI, Liquid Glass, and SwiftData, the app has zero network calls, no accounts, no analytics, and works fully offline in airplane mode. [![User Avatar](https://storage.googleapis.com/papyrus_images/9b6cfbeca2f824acf7093154ef80bc1d2e7485246db22713ff06a1421ef3a4cd.jpg)](https://twitter.com/elirousso) [Eli Rousso](https://twitter.com/elirousso) [@elirousso](https://twitter.com/elirousso) [](https://twitter.com/elirousso/status/2079594911637094442) ![⚪️](https://abs-0.twimg.com/emoji/v2/72x72/26aa.png) New App: Slate — a voice journal built on one rule: nothing leaves your phone. 100% Apple, top to bottom. Transcription by SpeechAnalyzer on device. Reflection by the 3 billion parameter Apple Intelligence model on the Neural Engine. Storage in SwiftData. Swift 6, SwiftUI, ![](https://storage.googleapis.com/papyrus_images/14790865be73da1fbafa93b34c54c4456b57683cb92f476cd75e3a7dc4d7486d.jpg) [6,074](https://twitter.com/elirousso/status/2079594911637094442)[ 3:50 PM • Jul 21, 2026 ](https://twitter.com/elirousso/status/2079594911637094442) My high-school classmate [Abhi Ingle,](https://www.linkedin.com/in/ingle-abhi/)  Chief Product and Strategy Officer at Sambanova in a conversation with [Diana Goovaerts](https://x.com/DiaMariesbeat) of [Fierce Network](https://x.com/FierceNetwork_) articulating the differentiation Sambanova brings to inference workloads [ Thank you Diana Goovaerts for the great conversation this week. Your Fierce Network article does a wonderful job contextualizing our conversation and making it relevant to a broad audience. As you... | Abhi I. ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Thank you Diana Goovaerts for the great conversation this week. Your Fierce Network article does a wonderful job contextualizing our conversation and making it relevant to a broad audience. As you put it so elegantly, we aren't trying to out muscle anyone, we just want to meet customers where they are - our focus is on making customers successful. https://www.linkedin.com ![Thank you Diana Goovaerts for the great conversation this week. Your Fierce Network article does a wonderful job contextualizing our conversation and making it relevant to a broad audience. As you... | Abhi I.](https://storage.googleapis.com/papyrus_images/d6502b27c83b35a40ae22dc5dcbf1bc6c9793adb5273c72e63d160ba6d31eda1.jpg) ](https://www.linkedin.com/feed/update/urn:li:activity:7485809678204145665/) [ SambaNova targets AI inference boom with chips built for existing data centers ------------------------------------------------------------------------------ SambaNova told Fierce Network its unique architecture can do the job faster, cooler and with less ener | SambaNova's air-cooled inferencing chips could be a huge deal for data center operators - and telcos - looking to maximize existing facilities. https://www.fierce-network.com ![SambaNova targets AI inference boom with chips built for existing data centers](https://storage.googleapis.com/papyrus_images/dc39487a48a219912ae1d10d214b8ef29079caa95d17b2a5b234d7830fe9afae.png) ](https://www.fierce-network.com/cloud/sambanova-targets-ai-inference-boom-chips-built-existing-data-centers) [Alex](https://t.me/zk_alex) asked if any had tried [ GitHub - Graphify-Labs/graphify: Turn any codebase, with its docs, SQL schemas, configs, and PDFs, into a queryable knowledge graph. A /graphify skill for Claude Code, Cursor, Codex, and Gemini CLI: local deterministic AST parsing, every edge explained, no vector store. ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ Turn any codebase, with its docs, SQL schemas, configs, and PDFs, into a queryable knowledge graph. A /graphify skill for Claude Code, Cursor, Codex, and Gemini CLI: local deterministic AST parsing, every edge explained, no vector store. - Graphify-Labs/graphify https://github.com ![GitHub - Graphify-Labs/graphify: Turn any codebase, with its docs, SQL schemas, configs, and PDFs, into a queryable knowledge graph. A /graphify skill for Claude Code, Cursor, Codex, and Gemini CLI: local deterministic AST parsing, every edge explained, no vector store.](https://storage.googleapis.com/papyrus_images/ce565717926e0d51522d57ea41a6a78edfe59e60f8c5ec3a0d167487233572a0.png) ](https://github.com/Graphify-Labs/graphify) via [Alex](https://t.me/zk_alex) Wild world of AI marketing. Doommaxxing has been the preferred Marketing strategy of Anthropic for years, OpenAI is just catching up. Remember when Amodei said GPT2 was too dangerous to release in 2019? Good blog on this (NYT is paywalled):... [ Dear AI Companies: Stop the "Doom Trolling" - Cal Newport --------------------------------------------------------- Imagine, for a moment, the following scenario. The Ford Motor Company releases a slick whitepaper making the alarming claim that they're concerned their popular F-150 ... Read more https://calnewport.com ![Dear AI Companies: Stop the "Doom Trolling" - Cal Newport](https://storage.googleapis.com/papyrus_images/a52c62a32923cf280d67e8b88592e38aebc1c02799a0cbd5d2ac3ea2d1c106c2.png) ](https://calnewport.com/dear-ai-companies-stop-the-doom-trolling/) via [James Chan](https://t.me/motochan) with a longish post on his experiments with Laguna S 2.1 though he subsequently mentioned that given updated weights were uploaded by poolside, he would be doing this again Some learnings from deploying Laguna S 2.1 NVFP4 on my GB10 (ignore project references; I had Kimi K3 and Fable check on Laguna’s work): **Read so far:** strong at reading the codebase (comprehension flawless, type-level reasoning dead-on), degrades on grounded execution — fabricates wiring/fields it can't see, over-claims in adversarial review. 4/5 carried a concrete defect only execution exposed. **Where Laguna is genuinely strong** •⁠  ⁠Code comprehension — flawless both projects (R-D1, C-D1 exact, incl. Dec→Jan tier math) •⁠  ⁠Pure-function implementation — C-D3 dropped in cleanly, verified by execution •⁠  ⁠Test authoring when the harness contract is spelled out — C-D4 perfect (17/17); R-D4 degraded because it had to infer the fixture shape itself and fabricated a field **Two recurring failure modes** — both confidently wrong 1.⁠ ⁠Judgment calls — correct trace, wrong verdict. C-D2: line-perfect root-cause, then declared an intended design an "off-by-one defect." R-D2: nailed the mechanism, but its fix inverted polarity (would've hidden discoverable members and kept leaking hidden ones). Security reviews find real issues but dilute them with invented ones — even ignoring an explicit "don't invent issues" instruction (C-D5). 2.⁠ ⁠Cross-boundary integration — fabricates wiring it can't see: [req.app](http://req.app).env.db (doesn't exist here), missing await on an async auth gate, non-existent test helpers. **Fable's sharpening of Kimi K3 thesis:** the dangerous part isn't that it errs — it's that errors arrive with full confidence and correct-looking reasoning attached. That's exactly what makes an ungated lane unsafe and a gated lane fine. **(interim) Bottom line (both Fable & Kimi K3 agree):** Laguna is a strong gated drafter, not an autonomous shipper. Let it draft pure functions + tests off a concrete in-repo reference; I own every judgment call, security verdict, and integration seam. The adversarial gate is load-bearing — R-D2 and the earlier live privacy leak would both have shipped unreviewed. **My custom benchmark script missed 2 areas that I continued to test for:** long-context, and negative-constraint following at length; following findings are post-those 2 catch-up tests. **Read:** trust Laguna to find & explain in big contexts — never trust its emitted file:line lists or hard structural-constraint adherence without an automated post-check. **The consolidated thesis, now evidence-backed across 13 tasks + two independent graders:** Laguna's reading is bimodal — trust its semantic comprehension (what the code means, where logic lives, needle retrieval at 144KB) considerably; never trust its mechanical bookkeeping (file:line citations, hard-constraint compliance) without an automated post-check. Fable's framing: the danger isn't that it errs — it's that errors arrive confident, with correct-looking reasoning attached. **Operating rule for the token-reduction pipeline:** 1.⁠ ⁠Let Laguna lead — comprehension, pure-function drafting, needle-in-haystack, test authoring off a concrete reference 2.⁠ ⁠Gate is load-bearing on: every reference sweep (grep-verify its site list — don't trust the citations), security verdicts, cross-boundary wiring, and any hard-constraint output (lint/post-check it) I haven't tried FP8 Laguna S2.1 across both my Asus Ascent GX10; not sure if it's supported, need to investigate but MoE so high chances it can Anthropic announces the public beta of the Claude Security plugin for Claude Code, enabling terminal-based scans of code changes before commits or full codebase reviews using existing Claude inference.  The tool applies AI reasoning to trace data flows across files, detect context-dependent vulnerabilities like injection flaws or auth bypasses, and uses adversarial self-verification to cut false positives while proposing style-preserving patches.  It runs entirely within the user's environment, integrates via webhooks to Slack or Jira, supports scheduled scans, and requires human review of all fixes for security-critical systems. [![User Avatar](https://storage.googleapis.com/papyrus_images/dc69ec28be72de3fb258b612f16a89ba4ffa6bf64d6b63683c69051b6c38c662.jpg)](https://twitter.com/claudeai) [Claude](https://twitter.com/claudeai) [@claudeai](https://twitter.com/claudeai) [](https://twitter.com/claudeai/status/2079990597973057691) The Claude Security plugin for Claude Code is now available in beta. Scan your changes for vulnerabilities before you commit, or run a full scan across your codebase, all from your terminal on the Claude inference you already run. ![](https://pbs.twimg.com/amplify_video_thumb/2079986661207732224/img/6P-2l2p-49hKFPco.jpg) [14.9K](https://twitter.com/claudeai/status/2079990597973057691)[ 6:02 PM • Jul 22, 2026 ](https://twitter.com/claudeai/status/2079990597973057691) [https://claude.com/product/claude-security#beta](https://claude.com/product/claude-security#beta) Reasonable buzz on with Chinese internet scena aka WeChat groups/channels about supposedly transcripts \[IMO, more likely quotes\] from the 4-hour investor briefing that DeepSeek founder (Liang Wenfeng) recently held Feel free to keep an healthy scepticism on the veracity of these statements though they seem to be inline with many people's perception of Deepseek's as well as the founders goals. Original (in Chinese, from Tencent Tech) [ 4小时、118个回答，梁文锋内部交流回应一切 ---------------------- 全程实录。 https://mp.weixin.qq.com ![4小时、118个回答，梁文锋内部交流回应一切](https://storage.googleapis.com/papyrus_images/3c88d2bc5a7f0890f2aa96188760156640a8e1a2ce7420a1effc8330b7e9abaa.jpg) ](https://mp.weixin.qq.com/s/Rqv-4yFGUTuLxMaRSt1XDg?chain_id=a_QaRK54YZcKON.1l640nl&global_content=%7B%22promote_id%22%3A13766%2C%22sub_promote_id%22%3A33%2C%22f%22%3A%22news.futunn.com%2Fen%2Fpost%2F76458063%2Fliang-wenfeng-s-four-hour-investor-meeting-transcript-4-hours%22%7D) English Translation [https://news.futunn.com/en/post/76458063/liang-wenfeng-s-four-hour-investor-meeting-transcript-4-hours?level=3&data\_ticket=1784819518986209](https://news.futunn.com/en/post/76458063/liang-wenfeng-s-four-hour-investor-meeting-transcript-4-hours?level=3&data_ticket=1784819518986209) Friday 24th July 2026 --------------------- Ling-3.0-flash is a 124B MoE model with 5.1B active parameters per token that matches or exceeds Ant Group's 1T flagship on benchmarks like SWE-Bench, Terminal-Bench, and agent tasks using far fewer resources.  It employs native hybrid-linear attention with KDA for long-range memory and MLA layers in 5:1 ratio, enabling 256K native context scalable to 1M tokens alongside efficient 1/64 expert activation  [![User Avatar](https://storage.googleapis.com/papyrus_images/f15adde7eb2681980bcf9473ca33a08a73c8618afe59323fa5338993bead1440.jpg)](https://twitter.com/AntLingAGI) [Ant Ling](https://twitter.com/AntLingAGI) [@AntLingAGI](https://twitter.com/AntLingAGI) [](https://twitter.com/AntLingAGI/status/2080351022028095681) Today, we’re releasing Ling-3.0-flash—a hybrid-reasoning MoE model built for production-scale agents. 124B parameters. Just 5.1B active per token. With 1/8 of the total and 1/12 of the active parameters, it matches or beats our 1T flagship model on most benchmarks shown. ![](https://storage.googleapis.com/papyrus_images/1fc54e45eb8b932283ba271dc0b5e3172fb6c938145a4a3d493c8744caadf7db.jpg) [1,751](https://twitter.com/AntLingAGI/status/2080351022028095681)[ 5:55 PM • Jul 23, 2026 ](https://twitter.com/AntLingAGI/status/2080351022028095681) [![User Avatar](https://storage.googleapis.com/papyrus_images/8246b16ae278183515d2b9295f484ef1ab242259358a34aa42e2579a5bb4ce21.jpg)](https://twitter.com/TeksEdge) [David Hendrickson](https://twitter.com/TeksEdge) [@TeksEdge](https://twitter.com/TeksEdge) [](https://twitter.com/TeksEdge/status/2080366619411501364) ![🤩](https://abs-0.twimg.com/emoji/v2/72x72/1f929.png) A 124B model activating only 5.1B parameters while approaching its 1T sibling’s performance. It's ~ as good as Sonnet-4.6 to run locally. The mystery model that appeared before its announcement is official. [@TheInclusionAI](https://twitter.com/TheInclusionAI) has officially released Ling-3.0-flash, a ![](https://storage.googleapis.com/papyrus_images/cfa96443733f3c7b07ccdbc2a77b30d0d455ec9a7a1b2b883dd7ea4a5f5f32f3.jpg) [![User Avatar](https://storage.googleapis.com/papyrus_images/f15adde7eb2681980bcf9473ca33a08a73c8618afe59323fa5338993bead1440.jpg)](https://twitter.com/AntLingAGI) [Ant Ling](https://twitter.com/AntLingAGI) [@AntLingAGI](https://twitter.com/AntLingAGI) [](https://twitter.com/AntLingAGI/status/2080351022028095681) Today, we’re releasing Ling-3.0-flash—a hybrid-reasoning MoE model built for production-scale agents. 124B parameters. Just 5.1B active per token. With 1/8 of the total and 1/12 of the active parameters, it matches or beats our 1T flagship model on most benchmarks shown. ![](https://storage.googleapis.com/papyrus_images/1fc54e45eb8b932283ba271dc0b5e3172fb6c938145a4a3d493c8744caadf7db.jpg) [477](https://twitter.com/TeksEdge/status/2080366619411501364)[ 6:57 PM • Jul 23, 2026 ](https://twitter.com/TeksEdge/status/2080366619411501364) [Command Code](https://commandcode.ai/) v1 recently released in stealth.  This is a model agnostic coding harness  similar to Opencode, Droid from Factory, Cline, Devin/Windsurf Cursor can also be considered as model agnostic but its upcoming ownership via xAI would make it a harness operated by a frontier lab similar to Codex/Claude Code The reason I'm highlighting this harness is that it has a US $1/month [plan](https://commandcode.ai/pricing) and also supports payments via [UPI](https://commandcode.ai/docs/resources/payment-methods) which is India's payment rails.  Therefore potentially useful to someone who would like model agnostic coding harness which is affordable relative to their purchasing power.   For some folks spending $20/month or even $200/month for a coding harness may not mean much, for some $10/month may mean a lot.   Feel free to share this to friends/family/acquaintainces in emerging markets if you think the price point would interest them [![User Avatar](https://storage.googleapis.com/papyrus_images/89a2f89eae7daddda81c8c850c915bc8f3eef15af2eb318eacfe774a684259af.jpg)](https://twitter.com/MrAhmadAwais) [Ahmad Awais](https://twitter.com/MrAhmadAwais) [@MrAhmadAwais](https://twitter.com/MrAhmadAwais) [](https://twitter.com/MrAhmadAwais/status/2080120000309088323) stealth released v1 of command code today. ![🙈](https://abs-0.twimg.com/emoji/v2/72x72/1f648.png) [85](https://twitter.com/MrAhmadAwais/status/2080120000309088323)[ 2:37 AM • Jul 23, 2026 ](https://twitter.com/MrAhmadAwais/status/2080120000309088323) [![User Avatar](https://storage.googleapis.com/papyrus_images/89a2f89eae7daddda81c8c850c915bc8f3eef15af2eb318eacfe774a684259af.jpg)](https://twitter.com/MrAhmadAwais) [Ahmad Awais](https://twitter.com/MrAhmadAwais) [@MrAhmadAwais](https://twitter.com/MrAhmadAwais) [](https://twitter.com/MrAhmadAwais/status/2080120861823209513) v1 is live, so much better than anything out there [ ![](https://storage.googleapis.com/papyrus_images/7832ffc54e32fb1db2746ca7c7c846c9931107cc7cb86a29270bb212fce19c71.jpg) commandcode.ai What's New in v1 | Command Code ------------------------------- Everything new from v0 to v1 of Command Code: the rewritten permission engine, background tasks, worktrees, sessions, context management, keybindings, mods, and more, with links to the full docs for... ](https://t.co/K5TlwkjBuf) [12](https://twitter.com/MrAhmadAwais/status/2080120861823209513)[ 2:40 AM • Jul 23, 2026 ](https://twitter.com/MrAhmadAwais/status/2080120861823209513) What's new in v1 [https://commandcode.ai/docs/whats-new](https://commandcode.ai/docs/whats-new) Not all the features originally slated for v1 mention in the tweet below are live as yet but seems like this is going to be a multi-surface harness \[CLI, TUI, GUI, SDK etc\] and will also be open source [![User Avatar](https://storage.googleapis.com/papyrus_images/89a2f89eae7daddda81c8c850c915bc8f3eef15af2eb318eacfe774a684259af.jpg)](https://twitter.com/MrAhmadAwais) [Ahmad Awais](https://twitter.com/MrAhmadAwais) [@MrAhmadAwais](https://twitter.com/MrAhmadAwais) [](https://twitter.com/MrAhmadAwais/status/2073493060370219438) Massive v1 rewrite of Command Code is underway. We’ve taken all the feedback from months, a couple of strong novel ideas, and built a strong harness core that is runtime and transport agnostic. This new harness will be an amazing kernel core for our existing CLI TUI as well as [139](https://twitter.com/MrAhmadAwais/status/2073493060370219438)[ 7:44 PM • Jul 4, 2026 ](https://twitter.com/MrAhmadAwais/status/2073493060370219438) [Etched](https://www.youtube.com/watch?v=BagWrgPww1o) raised $300M in Series C funding at a $10.3B valuation.  Investors: Sequoia, Andreessen Horowitz, Jane Street, Argo, SK Hynix. [![User Avatar](https://storage.googleapis.com/papyrus_images/1071f9ab56c8d26edb2e2121c5ae49d82c0199aac89efd9ff41ba6bb738cf6ec.jpg)](https://twitter.com/Etched) [Etched](https://twitter.com/Etched) [@Etched](https://twitter.com/Etched) [](https://twitter.com/Etched/status/2080307393699987849) We’ve raised $300M in Series C funding at a $10.3B valuation from Sequoia, Andreessen Horowitz, Jane Street, Argo, and SK Hynix. Our mission is to run the world's inference. This round accelerates production of our inference clusters. We've opened an 80,000-sqft, 10-MW facility [5,161](https://twitter.com/Etched/status/2080307393699987849)[ 3:01 PM • Jul 23, 2026 ](https://twitter.com/Etched/status/2080307393699987849) [https://x.com/stretchcloud/status/20803587855083196](https://x.com/stretchcloud/status/20803587855083196) [![User Avatar](https://storage.googleapis.com/papyrus_images/b53aea47a35f35e67e8b8111ad04890029df9368091b6a19ca545ba53ec59399.jpg)](https://twitter.com/victoralazarte) [victor lazarte](https://twitter.com/victoralazarte) [@victoralazarte](https://twitter.com/victoralazarte) [](https://twitter.com/victoralazarte/status/2080308224725582012) Etched is Diffusion's (and my) largest investment to date. I'm excited to be partnering with [@UbertiGavin](https://twitter.com/UbertiGavin), [@robertwachen](https://twitter.com/robertwachen), [@czhu1729](https://twitter.com/czhu1729), and the [@Etched](https://twitter.com/Etched) team. Three Harvard dropouts started Etched in 2022 believing, that "inference is going to be the biggest market in the world. [![User Avatar](https://storage.googleapis.com/papyrus_images/1071f9ab56c8d26edb2e2121c5ae49d82c0199aac89efd9ff41ba6bb738cf6ec.jpg)](https://twitter.com/Etched) [Etched](https://twitter.com/Etched) [@Etched](https://twitter.com/Etched) [](https://twitter.com/Etched/status/2080307393699987849) We’ve raised $300M in Series C funding at a $10.3B valuation from Sequoia, Andreessen Horowitz, Jane Street, Argo, and SK Hynix. Our mission is to run the world's inference. This round accelerates production of our inference clusters. We've opened an 80,000-sqft, 10-MW facility [210](https://twitter.com/victoralazarte/status/2080308224725582012)[ 3:05 PM • Jul 23, 2026 ](https://twitter.com/victoralazarte/status/2080308224725582012) Black Forest Labs announces FLUX 3, a single unified multi-modal model that generates images, videos, audio, and predicts actions for robotics, producing highly realistic outputs across styles  FLUX 3 Video is available in early access, demonstrated by a diverse montage of generated clips ranging from photorealistic ocean waves and breaking glass to surreal desert scenes and animated characters.  The model extends to physical AI via FLUX-mimic partnership with mimic robotics, enabling dexterous robot tasks like automotive assembly tested at Audi, with image samples showing photorealistic, illustrative, and cinematic capabilities. [![User Avatar](https://storage.googleapis.com/papyrus_images/48e40d0b60e6cc4b4d50c3a578891946885121c7657784e16afb443921212c93.jpg)](https://twitter.com/bfl_ai) [Black Forest Labs](https://twitter.com/bfl_ai) [@bfl\_ai](https://twitter.com/bfl_ai) [](https://twitter.com/bfl_ai/status/2080308988961554582) Introducing FLUX 3. One multi-modal model for Image, Video, Audio and Action-Prediction. Creations are truer to life in every kind of style. FLUX 3 Video is now available in early access (link below). Jointly trained in one unified architecture, our model can be extended to ![](https://pbs.twimg.com/amplify_video_thumb/2080308957898481664/img/6WEYl2biM9RO6YtB.jpg) [5,793](https://twitter.com/bfl_ai/status/2080308988961554582)[ 3:08 PM • Jul 23, 2026 ](https://twitter.com/bfl_ai/status/2080308988961554582) Celeris Labs launched celeris-1, a general-purpose LLM using a hybrid autoregressive-diffusion inference architecture that achieves 157ms p50 response latency and 1,280 tokens per second throughput.  Benchmarks position it at 76% MMLU-Pro accuracy, placing it near GPT-5-mini (78%) and GPT-5 (81%) on a speed-vs-intelligence chart while delivering 15-17x faster responses than those models.  The model is available immediately via sign-up at [celeris.ai](http://celeris.ai), advancing the lab's goal of maximizing useful intelligence per unit time through diffusion decoding innovations.  Pricing is $2/M for input and $6/M for output.  However context-window is a hard 8,192 token limit.  Only available in the US for now  [![User Avatar](https://storage.googleapis.com/papyrus_images/119d9457bf365668a0b011c6f7f8efd15a0e35c1b34903b59ef7d5f74870d589.jpg)](https://twitter.com/Celeris_ai) [Celeris Labs](https://twitter.com/Celeris_ai) [@Celeris\_ai](https://twitter.com/Celeris_ai) [](https://twitter.com/Celeris_ai/status/2080442996403933630) Introducing celeris-1. A general purpose language model which delivers near-GPT-5 level intelligence with 15x faster response times. We’ve developed a new inference architecture that uses diffusion techniques instead of conventional autoregressive generation - unlocking ![](https://storage.googleapis.com/papyrus_images/2d05922a3e69715131f91c5e188bc19df41bd77b1d36fd35c435de0f1cd9553b.jpg) [945](https://twitter.com/Celeris_ai/status/2080442996403933630)[ 12:00 AM • Jul 24, 2026 ](https://twitter.com/Celeris_ai/status/2080442996403933630) [![User Avatar](https://storage.googleapis.com/papyrus_images/6525dcc5bb5883b93bd795e2c734fb94fa79cd54b506d55f2586a96713784005.jpg)](https://twitter.com/tom_w_hamer) [Tom Hamer](https://twitter.com/tom_w_hamer) [@tom\_w\_hamer](https://twitter.com/tom_w_hamer) [](https://twitter.com/tom_w_hamer/status/2080443293582987643) Excited to launch Celeris-1 - a general purpose language model which delivers near-GPT-5 level intelligence with 15x faster response times. Celeris-1 delivers a p50 response latency of 157ms - around 15× faster than GPT-5-mini and 17× faster than GPT-5 - while scoring comparably[ ![](https://storage.googleapis.com/papyrus_images/61f3eb40e4f52ce7695d21a97cfcc2d4b27215028eac034f308e91d9f4dee797.jpg) celeris.ai Celeris - An AI research lab building the world's fastest LLM ------------------------------------------------------------- Celeris is an AI research lab building the world's fastest LLM. Frontier-level reasoning at real-time latency. ](https://t.co/WXZWiTHPcO) [![User Avatar](https://storage.googleapis.com/papyrus_images/119d9457bf365668a0b011c6f7f8efd15a0e35c1b34903b59ef7d5f74870d589.jpg)](https://twitter.com/Celeris_ai) [Celeris Labs](https://twitter.com/Celeris_ai) [@Celeris\_ai](https://twitter.com/Celeris_ai) [](https://twitter.com/Celeris_ai/status/2080442996403933630) Introducing celeris-1. A general purpose language model which delivers near-GPT-5 level intelligence with 15x faster response times. We’ve developed a new inference architecture that uses diffusion techniques instead of conventional autoregressive generation - unlocking ![](https://storage.googleapis.com/papyrus_images/2d05922a3e69715131f91c5e188bc19df41bd77b1d36fd35c435de0f1cd9553b.jpg) [763](https://twitter.com/tom_w_hamer/status/2080443293582987643)[ 12:01 AM • Jul 24, 2026 ](https://twitter.com/tom_w_hamer/status/2080443293582987643) [https://docs.celeris.ai/models](https://docs.celeris.ai/models) [https://docs.celeris.ai/availability](https://docs.celeris.ai/availability) Jensen Huang's first X post shares a July 24, 2026 open letter signed by NVIDIA and major tech firms advocating open-weight AI models to advance American AI leadership.  The letter highlights how open models strengthen cybersecurity, accelerate innovation and competition, expand AI access across industries, and support national sovereignty alongside closed frontier models.  It compares open AI development to the 1980s open-source software movement and calls on policymakers to foster a pluralistic ecosystem without premature restrictions on open models. [![User Avatar](https://storage.googleapis.com/papyrus_images/3629326653742c9f29e842b70703bbc9654b316a9ffbc00099083c0ebfdbca0a.jpg)](https://twitter.com/JensenHuang) [Jensen Huang](https://twitter.com/JensenHuang) [@JensenHuang](https://twitter.com/JensenHuang) [](https://twitter.com/JensenHuang/status/2080643682408321103) For my first post, I’m sharing a letter @NVIDIA signed on why open models matter. AI will transform every industry, power every company, and be built by every country. Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. ![](https://storage.googleapis.com/papyrus_images/f7913b972041aa6115d22b62662c691be2b44d0a891be5d619c72a37fb8ca615.jpg) [114.9K](https://twitter.com/JensenHuang/status/2080643682408321103)[ 1:18 PM • Jul 24, 2026 ](https://twitter.com/JensenHuang/status/2080643682408321103) Registration is now open for AIME 2026 — AI Micro-Drama Europe, a European competition dedicated to AI-generated vertical micro-dramas. Co-organized by the European Microdrama Association and French micro-drama label MINI CLAP, AIME offers creators an opportunity to experiment with a rapidly emerging form of mobile-first storytelling—and to connect their work with production support, industry opportunities and international audiences. You can register now and submit your completed work later. No finished video is required at the initial registration stage. Registration deadline: July 31, 2026, at 23:59:59 Paris time Submission deadline: August 31, 2026, at 23:59:59 Paris time [https://www.linkedin.com/feed/update/urn:li:activity:7486329654782373888/](https://www.linkedin.com/feed/update/urn:li:activity:7486329654782373888/) Saturday 25th July 2026 ----------------------- Anthropic Launches Claude Opus 5 with Frontier Performance at Half the Cost Available now on paid Claude plans, Claude Code, and the API, Opus 5 prices at $5 per million input tokens and $25 per million output tokens, the same as its predecessor. It leads benchmarks in agentic coding, business workflows, and novel problem-solving like ARC-AGI-3 at 30.2%, with independent tests confirming its top Intelligence Index score. Developers praise the value and speed, though some note it needs fresh prompts to shine, positioning it as an efficient alternative to pricier models. [![User Avatar](https://storage.googleapis.com/papyrus_images/e6f736ef39a919cd51cb27919ba706916ba8718627b6635dde46a15532a2c415.png)](https://twitter.com/ClaudeDevs) [ClaudeDevs](https://twitter.com/ClaudeDevs) [@ClaudeDevs](https://twitter.com/ClaudeDevs) [](https://twitter.com/ClaudeDevs/status/2080703243722854516) Opus 5 is live in Claude Code and the Claude Platform today. A few things worth knowing: ![🧵](https://abs-0.twimg.com/emoji/v2/72x72/1f9f5.png) [x.com/claudeai/statu…](https://t.co/jAdUagXt9C) [![User Avatar](https://storage.googleapis.com/papyrus_images/dc69ec28be72de3fb258b612f16a89ba4ffa6bf64d6b63683c69051b6c38c662.jpg)](https://twitter.com/claudeai) [Claude](https://twitter.com/claudeai) [@claudeai](https://twitter.com/claudeai) [](https://twitter.com/claudeai/status/2080699495453528290) Introducing Claude Opus 5. It's a thoughtful and proactive model that comes close to the frontier intelligence of Fable 5 at half the price. ![](https://pbs.twimg.com/media/HOAifjuWcAESshR.jpg) [5,560](https://twitter.com/ClaudeDevs/status/2080703243722854516)[ 5:14 PM • Jul 24, 2026 ](https://twitter.com/ClaudeDevs/status/2080703243722854516) [![User Avatar](https://storage.googleapis.com/papyrus_images/3feb1068d380a616edb9063ba3b1f774758f970c1f256e3208da333d8fa27575.jpg)](https://twitter.com/trq212) [Thariq](https://twitter.com/trq212) [@trq212](https://twitter.com/trq212) [](https://twitter.com/trq212/status/2080710971228918066) We removed ~80% of the Claude Code system prompt for our newest models, this is what we've learned about writing system prompts, skills and Claude.MDs for them. [x.com/i/article/2080…](https://t.co/6DZwSrZjE9) [7,052](https://twitter.com/trq212/status/2080710971228918066)[ 5:45 PM • Jul 24, 2026 ](https://twitter.com/trq212/status/2080710971228918066) [![User Avatar](https://storage.googleapis.com/papyrus_images/9a24e89bcea6545e47519412d53c6a01304c5edd527986785aa7374097b8e18a.jpg)](https://twitter.com/danshipper) [Dan Shipper 📧](https://twitter.com/danshipper) [@danshipper](https://twitter.com/danshipper) [](https://twitter.com/danshipper/status/2080700057892815114) BREAKING: Claude Opus 5 is OUT NOW! And…it’s a hard model to love. We’ve spent the last week [@every](https://twitter.com/every) testing it across coding, writing, knowledge work, and our internal agent. It argued with instructions, stopped before the work was finished, and generally didn’t play well ![](https://pbs.twimg.com/amplify_video_thumb/2080699905501274112/img/9FVMuLdML5uwgZyt.jpg) [![User Avatar](https://storage.googleapis.com/papyrus_images/dc69ec28be72de3fb258b612f16a89ba4ffa6bf64d6b63683c69051b6c38c662.jpg)](https://twitter.com/claudeai) [Claude](https://twitter.com/claudeai) [@claudeai](https://twitter.com/claudeai) [](https://twitter.com/claudeai/status/2080699495453528290) Introducing Claude Opus 5. It's a thoughtful and proactive model that comes close to the frontier intelligence of Fable 5 at half the price. ![](https://pbs.twimg.com/media/HOAifjuWcAESshR.jpg) [2,014](https://twitter.com/danshipper/status/2080700057892815114)[ 5:02 PM • Jul 24, 2026 ](https://twitter.com/danshipper/status/2080700057892815114) [ What's new in Claude Opus 5 --------------------------- Overview of new features and behavior changes in Claude Opus 5. https://platform.claude.com ![What's new in Claude Opus 5](https://storage.googleapis.com/papyrus_images/a2ba989463949a7aa139030064e134b98b1f375e44699c79e466be05f41307f4.png) ](https://platform.claude.com/docs/en/about-claude/models/whats-new-opus-5) The single biggest theme: **Opus 5 is so much more capable that most old prompting "best practices" are now counterproductive.** Anthropic's own Claude Code team removed **~80% of their system prompt** for Opus 5 and Fable 5 with no loss on coding evals. The bottleneck is no longer the model — it's over-constraining it. **Official prompting guide:** [https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5)   **Context engineering blog:** [https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models](https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models)   **Auto-fix tool:** Run /doctor in Claude Code to automatically rightsize your [CLAUDE.md](http://CLAUDE.md) and skills. * * * Below is my personal website which aggregates links to many of my socials as well as the various content and community that I curate. Feel free to share this link to others who you think may find this content/community useful to them [**https://linktr.ee/goolamabbas**](https://linktr.ee/goolamabbas) The cover image of this newsletter via generated via the Seedream 5.0 Pro model within the [**Krea**](https://www.krea.ai/refer/EJQQP9FJ) tool via the following prompt ![](https://paragraph.com/editor/callout/information-icon.png) A surrealist, dreamlike masterpiece of a lush, misty green hillside enveloped in thick, ethereal fog. The aesthetic is defined by a vintage 2000s CCD digicam texture, featuring characteristic digital grain, slight motion blur, and a soft-focus glow. Low-contrast, dreamy pastel color palette with muted greens and hazy whites. The lighting is diffused and atmospheric, as if captured during a foggy dawn. Masterful composition, ethereal and calm, high-art photography style, 4k, cinematic atmosphere, nostalgic filmic quality. --- *Originally published on [This Week in All Things AI](https://paragraph.com/@twiata/this-week-in-all-things-ai-week-30-2026)*