# This Week in All Things AI - Week 10-2026

> Sunday 1st March 2026 to Saturday 7th March 2026

**Published by:** [This Week in All Things AI](https://paragraph.com/@twiata/)
**Published on:** 2026-03-08
**URL:** https://paragraph.com/@twiata/this-week-in-all-things-ai-week-10-2026

## Content

The frontier labs are relentless in their launch of new models with Google Deepmind launching Gemini-3.1 Flash-Lite and OpenAI launching GPT-5.4 with impressive performance for their class. Cognition which makes Devin and Windsurf showed an early preview of their model SWE-1.6 which has a huge jump over their previous SWE-1.5 and offers 950 tokens/second inference via their Cerebras partnership Anthropic's new feature of importing preferences/context from other AI providers into Claude is its salvo to be more attractive to consumers with continuous improvements to Claude Code as well as Cowork and also taking advantage of more organisations offering skills in Anthropic format such as Dune, Coinmarketcap Readers who think others in their family, friends and acquaintances who are curious in knowing more about rapidly evolving AI tools/services/use cases and would benefit from being subscribers of this weekly newsletter are encouraged to share this publication link to them and invite them to subscribe The following messages were posted on the 'All Things AI' Telegram group from Sunday 1st Mar 2026 to Saturday 7th Mar 2026Sunday 1st March 2026Anthropic created a process to bring over preferences and context from other AI providers to Claude For those who are using other providers and fear starting from scratch with Claude, recommend going through the below landing page and read it completely to understand the process https://claude.com/import-memoryMonday 2nd March 20261Cognition Unveils SWE-1.6 with 55.8% SWE-Bench Pro Score The new model jumps from SWE-1.5's 40.1% on the benchmark, thanks to reinforcement learning that used 100 times more compute on thousands of NVIDIA GB200 chips—while keeping its speedy 950 tokens per second. It edges out open-source rivals and matches top closed models like Claude Opus 4.6, with internal tests showing solve rates climb from 52.4% to 68.7% on engineering tasks. Tweaks are planned for quirks like overthinking. Team members shared their excitement, with CEO Scott Wu calling it a big milestone and researchers praising the 'insane team' effort Cognition @cognition We are sharing an early preview of our ongoing SWE-1.6 training run. It significantly improves upon SWE-1.5 while being post-trained on the same pre-trained model - and it runs equally as fast at 950 tok/s. On SWE-Bench Pro it exceeds top open-source models. The preview model 1,194 5:42 AM • Mar 2, 2026 Scott Wu @ScottWu46 Early days but a big milestone for us! This model is still in preview and we expect to tune behavior a lot over the coming weeks - but wanted to get folks a snapshot as soon as we could. Cognition @cognition We are sharing an early preview of our ongoing SWE-1.6 training run. It significantly improves upon SWE-1.5 while being post-trained on the same pre-trained model - and it runs equally as fast at 950 tok/s. On SWE-Bench Pro it exceeds top open-source models. The preview model 246 6:34 AM • Mar 2, 2026 Silas Alberti @silasalberti Over the last few months we started building our research team at Cognition and we've come a long way! It's been exciting to figure out what it takes to build a large-scale post-training stack from scratch and push towards the frontier. My personal take is it's been easier than Cognition @cognition We are sharing an early preview of our ongoing SWE-1.6 training run. It significantly improves upon SWE-1.5 while being post-trained on the same pre-trained model - and it runs equally as fast at 950 tok/s. On SWE-Bench Pro it exceeds top open-source models. The preview model 222 6:13 AM • Mar 2, 2026 nader dabit @dabit3 Impressive results so far from SWE-1.6, and at 950 tokens/s it doesn't sacrifice speed for intelligence. Now rolling out early access to a subset of users in Windsurf. Cognition @cognition We are sharing an early preview of our ongoing SWE-1.6 training run. It significantly improves upon SWE-1.5 while being post-trained on the same pre-trained model - and it runs equally as fast at 950 tok/s. On SWE-Bench Pro it exceeds top open-source models. The preview model 57 6:51 AM • Mar 2, 2026 2Paul Graham, the OG of Y Combinator hinting at the much improved capabilities of Replit v4 Paul Graham @paulg Amjad showed me Replit's latest stuff. They're about to redefine vibe coding in a way that will seem obvious in retrospect. A lot of the biggest ideas have that quality. 3,496 1:33 PM • Mar 2, 2026 3via Yat Siu had an opportunity to tell Peter that crypto isn't bad for AI at all, on stage he actually said "I don't hate crypto" but do think the scammers out there scarred him very badly Yat Siu @ysiu 1/ Had the great pleasure to speak at @imperialisoc @imperialcollege on the importance of AI & crypto followed by a great talk and then panel with @steipete @simonsquibb and witnessing the incredible excitement and builder energy! The main points I made Robby Yung ⦿⦿⦿ @viewfromhk @ysiu explaining why crypto and the agentic web were meant for each other at @imperialaisoc 92 4:58 AM • Mar 2, 2026 4via Simon Davis I created a guide to give your agent ultimate context for free by automatically hooking up your OpenClaw to notes from every meeting and discussion. Hope some of you find this useful. via Marc McGinley Dune | We Are Hiring! @Dune Dune MCP is live Plug Dune directly into @claudeai, @ChatGPTapp, @cursor_ai, and more. Search tables. Write queries. Build charts. Check Usage. All from a single prompt. Your AI just became a Dune power user. 1,122 10:15 PM • Mar 2, 2026 Tuesday 3rd March 20261Chintan Turakhia, Senior Director of Engineering at Coinbase, in conversation with Clairo Vo, founder of ChatPRD and host of the 'How I AI' podcast Chintan led the transformation of a 1,000-plus-engineer organization to embrace AI tools at scale. When tasked with rewriting Coinbase’s self-custody wallet into a consumer social app in just six to nine months, Chintan turned to AI as a force multiplier claire vo 🖤 @clairevo Sure, you can vibe code but have you ever shipped so much with AI you literally break GitHub? That’s what @chintanturakhia and the team at @coinbase did as they pushed the edge of engineering with AI. This week, Chintan and I chat about how to get 1000s of engineers cooking 160 9:53 PM • Mar 2, 2026 2via Tom Ho MCP, x402, claude code & Openclaw skills for CoinMarketCap CoinMarketCap @CoinMarketCap AI agents are getting smarter, but they still need market context. Today, we’re launching 4 AI Agent-focused products: MCP for real-time data x402 support for CoinMarketCap APIs Skills for Claude Code Skills for @openclaw Equip your AI agents with real-time 1,410 1:02 AM • Mar 3, 2026 3Anthropic Rolls Out Voice Mode for Claude Code Developers activate it via '/voice' and hold the spacebar to speak, streaming transcripts right into their code editor without overwriting text. The feature shines for quick ideas, refactoring tasks, and accessibility Thariq @trq212 Voice mode is rolling out now in Claude Code. It’s live for ~5% of users today, and will be ramping through the coming weeks. You'll see a note on the welcome screen once you have access. /voice to toggle it on! 17.2K 8:28 AM • Mar 3, 2026 4Singapore Offers Free AI Premium Access for Job Training Starting mid-2026, Singaporeans aged 25 and above can access premium subscriptions from Google, Manus, Microsoft, and OpenAI when enrolling in selected SkillsFuture courses. Manpower Minister Tan See Leng emphasized hands-on practice to help everyone adapt to AI changes in jobs, targeting 100,000 AI-savvy workers by 2029. The plan responds to calls for broader access amid regional skills gaps,Free premium AI subscriptions for those taking certain SkillsFuture courses from 2nd half of 2026The Government has been engaging providers such as Google, Manus, Microsoft and Open AI. Read more at straitstimes.com.https://www.straitstimes.com5New venture SecretSauce from my friends Simon Davis and Benjamin Chevalier Small blurb from the Forbes article which introduces their product ===== SecretSauce is designed to address what Davis calls the “brand memory” gap in AI. Rather than prompting from scratch each time, users upload brand assets or share a website. The system builds what the company calls a “codex”, a memory layer that encodes visual identity, tone and product rules. ==== via Simon Davis SecretSauce encodes your brand once and uses that intelligence to produce on-brand content by default. No prompting, no fixing, no starting from zero every time. Built on a system originally developed for live games at massive scale, it gets smarter with every interaction. The longer you use it, the better it knows your brand, and the harder it becomes to replicate with anything else. I'll make sure we bump anyone from this group for the free beta. You can sign up here.Wednesday 4th March 20261via Tom Ho Got Qwen 3.5 to run on iphone locally, testing on simulator for now nftom.eth @NFTom_ETH I built an iOS app to get Qwen3.5-2B model running locally on iPhone. Results aren't perfect, but it's working quite well. I even added a browser tool to search the web and Qwen knows when to use it. Free yourself from paying model providers or sending data out! 0 1:28 AM • Mar 4, 2026 2From Google Announcing Gemini 3.1 Flash-Lite! ⚡️ Our fastest and most cost-efficient Gemini 3 series model yet. A 45 % increase in output speed and it outperforms 2.5 Flash. It also has dynamic thinking levels to match task complexity. Google AI @GoogleAI Smarter. Faster. Gemini 3.1 Flash-Lite is here The model offers uncompromising speed & intelligence at scale by focusing on: — Cost-efficiency: Priced at just $0.25/1M input and $1.50/1M output tokens, it gets work done faster at a fraction of the cost of larger models, 1,734 12:41 AM • Mar 4, 2026 https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-flash-lite/?linkId=593831043check out the updated skill-creator with built-in support for test generation (e.g., to measure + optimize tricky things like skill trigger rate). available in Claude Code as plugin, https://claude.ai and Cowork. Lance Martin @RLanceMartin check out the updated skill-creator. i esp like built-in support for test generation (e.g., to measure + optimize tricky things like skill trigger rate). available in Claude Code as plugin, Claude.ai, + Cowork. 1,690 2:31 AM • Mar 4, 2026 Ryan Whitehead @ryan_whitehead We shipped evals and benchmarking in a skill-creator update today. Write tests, A/B compare skill versions, track regressions. No code required. Whether you're a dev or an SME, you now have real tools to validate that your skills work. claude.com/blog/improving… 1 3:08 AM • Mar 4, 2026 4via Coop Michael Truell @mntruell We believe Cursor discovered a novel solution to Problem Six of the First Proof challenge, a set of math research problems that approximate the work of Stanford, MIT, Berkeley academics. Cursor's solution yields stronger results than the official, human-written solution. 8,254 2:39 AM • Mar 4, 2026 5 Nils @broodsugar x.com/i/article/2023… 765 3:35 PM • Mar 4, 2026 Thursday 5th March 20261via Ben Jammin If anyone uses OpenClaw, here's a Morning Briefing use case of mine that you may be interested in It scrapes a bunch of different sources for me and compiles everything for me into bite sized info that I can read quickly: - My Newsletters - Youtube competitor videos - Product hunt launches - Hot Reddit Topics I'd be happy to help anyone set this kind of stuff up 2Tanishq Kumar, Tri Dao, and Avner May from Together Compute introduce Speculative Speculative Decoding (SSD) for up to 2x faster LLM inference Researchers Tanishq Kumar, Tri Dao, and Avner May from Together Compute released SSD, a new LLM inference algorithm. SSD speculates verification outcomes in parallel to enable asynchronous drafting and verification, eliminating overhead from the small draft model in traditional speculative decoding by preemptively predicting and preparing likely verification paths. The method achieves up to 2x speedup over the strongest existing inference engines. Tanishq Kumar @tanishqkumar07 I've been working on a new LLM inference algorithm. It's called Speculative Speculative Decoding (SSD) and it's up to 2x faster than the strongest inference engines in the world. Collab w/ @tri_dao @avnermay. Details in thread. 3,952 1:42 AM • Mar 5, 2026 Tri Dao @tri_dao Attack of the asynchronous machines. We’ve seen this a lot in GPU kernels. This time the same principle applies in speculative decoding Tanishq Kumar @tanishqkumar07 I've been working on a new LLM inference algorithm. It's called Speculative Speculative Decoding (SSD) and it's up to 2x faster than the strongest inference engines in the world. Collab w/ @tri_dao @avnermay. Details in thread. 521 3:09 AM • Mar 5, 2026 Avner May @avnermay Excited to announce our new LLM inference algorithm, speculative speculative decoding (SSD)! It is fast — up to 2x faster than state-of-the-art inference engines (vLLM, SGLang). Working on this with @tanishqkumar07 and @tri_dao was a blast. Details in thread: Tanishq Kumar @tanishqkumar07 I've been working on a new LLM inference algorithm. It's called Speculative Speculative Decoding (SSD) and it's up to 2x faster than the strongest inference engines in the world. Collab w/ @tri_dao @avnermay. Details in thread. 668 1:45 AM • Mar 5, 2026 3In the leadup to the highly awaited Replit Agent v4 release, Amjad Masad and Replit release a ~53 min documentary of the 'behind the scenes' leadup to Agent v3 which launched on Sep 21st 2025 Amjad Masad @amasad AI is compressing how we build. Roles collapse, roadmaps expire quickly, and you end up rewriting the product every few months. So we thought we’d give people a behind-the-scenes look. 21 Days to Launch, a Replit documentary. 817 1:45 AM • Mar 5, 2026 Friday 6th March 20261OpenAI Launches GPT-5.4 as Top Model for Professional Tasks The new GPT-5.4 rolls out immediately via API as gpt-5.4 and gpt-5.4-pro, plus in Codex and gradually to ChatGPT users on Plus, Team, Pro, Enterprise, and Edu plans. It shines in agentic tasks with native computer use—interpreting screenshots, generating code, and controlling mouse or keyboard across apps Benchmarks show it leading rivals like Claude 4.6 and Gemini 3.1 Pro, with testers like Matt Shumer calling it the world's best and coding 'essentially solved.' OpenAI @OpenAI GPT-5.4 Thinking and GPT-5.4 Pro are rolling out now in ChatGPT. GPT-5.4 is also now available in the API and Codex. GPT-5.4 brings our advances in reasoning, coding, and agentic workflows into one frontier model. 21.9K 2:10 AM • Mar 6, 2026 OpenAI Developers @OpenAIDevs GPT-5.4 is here. Native computer-use capabilities. Up to 1M tokens of context in Codex and the API. Best-in-class agentic coding for complex tasks. Scalable tool search across larger ecosystems. More efficient reasoning for long, tool-heavy workflows. openai.com/index/introduc… 6,166 2:12 AM • Mar 6, 2026 Matt Shumer @mattshumer_ I've been testing GPT-5.4 for the last week. In short, it is the best model in the world, by far. It's so good that it's the first model that makes the “which model should I use?” conversation feel almost over. The biggest surprise: I barely use Pro anymore! If you know me, 2,788 2:10 AM • Mar 6, 2026 2Netflix has acquired interpositive, a start-up founded by Ben Affleck that makes AI-powered tools for filmmakers. The system builds AI models from a film’s dailies to assist with postproduction tasks like color, relighting and VFX while keeping filmmakers “at the center of the process.” Bela Bajaria, Netflix’s CCO, says the tech will provide creatives “more choices, more control and more protection for their vision.”Netflix Acquires AI Filmmaking Start-Up Founded by Ben Affleck, Who Will Serve as Adviser to StreamerIn a rare acquisition, Netflix has bought InterPositive, a start-up founded by Ben Affleck that makes AI-powered tools for filmmakers.https://variety.comBen Affleck has also been active in talk-shows and media appearances articulating his view on AI in the context of film making 3via Robby Yung Hasan Toor @hasantoxr BREAKING: Someone just open sourced the missing layer for AI agents and it's genuinely insane. It's called LangWatch. The complete platform for LLM evaluation and AI agent testing trace, evaluate, simulate, and monitor your agents end-to-end before a single user sees them. 708 8:32 PM • Mar 4, 2026 4via Ben Jammin Anthropic Report Labor market impacts of AI: A new measure and early evidence https://cdn.sanity.io/files/4zrzovbb/website/dc7bcd0224644fce97cecb7f9e68dcd8434b35f1.pdf5approx hour long conversation between Lisa Huang creator of Gemini Gems alongwith Aakash Gupta ===== Gemini Gems, Claude Projects, custom GPTs. If you're not using any of them, you're working harder than you need to. The creator of Gemini Gems walked me through her entire setup: Aakash Gupta @aakashgupta Gemini Gems, Claude Projects, custom GPTs. If you're not using any of them, you're working harder than you need to. The creator of Gemini Gems walked me through her entire setup: 3:52 - The 3 Gems everyone needs 6:05 - Building a custom Gem 32:22 - Measuring your setup 125 8:07 AM • Mar 6, 2026 Saturday 7th March 20261via Fazri Zubair Just started using 5.4 Extra High and evaluating its ability in code generation. We'll keep you guys posted. Anyone have some early notes or results?2response to the above from Coop The first few tasks I was thinking it was better than Opus, but after giving it some more full features it didn’t beat Opus for me. I will test it with e2e testing with Playwrite / Stagehand this week as I hear that’s where it excels besides frontend design which I also haven’t tested yet. My code base is 10m LOC though so I am expecting a high bar to 5.4m compared to most.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 The cover image of this newsletter via generated via the Nano Banana 2 model within the Freepik tool via the following promptA futuristic stadium in the year 2056, surrounded by grassy fields and crowds walking on bridges to reach it. The building is white with black details. In front, you can see an ocean with some boats around it. A huge bridge leads from one side over that river to another island where other buildings stand. High resolution, hyper-realistic, high detail, sharp focus, depth of field, volumetric lighting, global illumination.

## Publication Information

- [This Week in All Things AI](https://paragraph.com/@twiata/): Publication homepage
- [All Posts](https://paragraph.com/@twiata/): More posts from this publication
- [RSS Feed](https://api.paragraph.com/blogs/rss/@twiata): Subscribe to updates
- [Twitter](https://twitter.com/yusufg): Follow on Twitter