This Week in All Things AI - Week 12-2026
The pace at which LLM's are being refreshed is insane. Week 12 had OpenAI's GPT-5.4-mini, Minimax-2.7 which helped design and optimize its own training and evaluation harnesses. Cursor's Composer-2 which itself was built on top of Kimi-K2.5. Anthropic has also been on a roll with its shipping velocity on Claude Cowork as well as Claude Code.
Other interesting things were Fal's release of its MCP server and Packy McCormick's and Pim de Witte's 19K word article on World models.
The following messages were posted on the ' All Things AI ' Telegram group from Sunday 15th Mar 2026 to Saturday 21st Mar 2026
Sunday 15th March 2026
| 1 | via Yat Siu MiroFish: The God View Engine A developer in China built an AI engine that generates thousands of digital humans - each with their own personality, memory, and behavior - drops them into a virtual world, and watches them predict the future. BuBBliK @k1rallik x.com/i/article/2032… 3,539 3:11 PM • Mar 14, 2026 Response from Jing This is super cool Wrote a summary for it here cookies (🍪,🍪) | 饼妹 @jinglingcookies this is actually insane, one undergrad just vibe-built a future-predicting sim engine solo |
be 20-year-old undergrad in China
obsessed with AI agents and graph computing
build BettaFish, a multi-agent opinion analyzer
goes viral
hits #1 GitHub trending, 20k GitHub
BuBBliK
@k1rallik
x.com/i/article/2032…
144 9:05 PM • Mar 14, 2026 | | 2 | Off-peak hours for Anthropic seem to coincide with Asia daylight hours.
2x worth of usage on various Claude services for the next 2 weeks.
Maybe future pricing of models will also be time-based
Aakash Gupta
@aakashgupta
This two-week promo tells you more about AI infrastructure economics than any earnings call this quarter.
Anthropic hit $19 billion in annualized revenue as of early March. They raised $30 billion at a $380 billion valuation five weeks ago. Claude Code alone crossed $2.5 billion
Claude
@claudeai
A small thank you to everyone using Claude: We’re doubling usage outside our peak hours for the next two weeks.
257 4:05 AM • Mar 15, 2026 |
Monday 16th March 2026
| 1 | via Robby Yung Christos Tzamos @ChristosTzamos 1/4 LLMs solve research grade math problems but struggle with basic calculations. We bridge this gap by turning them to computers. We built a computer INSIDE a transformer that can run programs for millions of steps in seconds solving even the hardest Sudokus with 100% accuracy 5,939 9:30 PM • Mar 11, 2026 Reply via Alex I dont get this "in model execution" just feels like the wrong direction. Why delegate to a stochastical model what a deterministic algorithm can solve infinitely more efficiently? The author managed to get a model to solve a sudoko in 3 minutes by burning >9m tokens, while it is trivially easy to write an algorithm that brute forces a sudoku in a fraction of a second. lol |
| 2 | Some members of this group may be aware that I curate a Notion page named 'AI/ML Resource Hub' though many a times I've thought of just removing it given how powerful answer engines have become. There is still some opinionated curation which may not resonate with everyone Recently whilst sampling friends/colleagues/acquaintainces on how they were providing context to their coding harness and researching some new tooling, I wrote the following new sub-page and incorporated that into the hub. Feedback greatly appreciated [Public: Nia versus Context7 | Notion Background about Nia via this tweet thread which first came to my attention as an alternative to Context7 https://yusuf-goolamabbas-53.notion.site |
| 3 | via Tim Happy to share this big x402/agent update courtest the Agent0 SDK team: Tim Cotten @CottenIO This is incredibly important. This is OUR tool for the agent economy. Agent0 SDK was built by members of the same team, like @marco_derossi, who led the ERC-8004 standard working group. As of this morning it brings native support for x402 payments into the same package that agent0 Lab @agent0lab Your agent just got an identity, a voice, and a wallet in one SDK. 8004 for identity. A2A for communication. x402 for payments. Live now on agent0 SDK. Here's how it works ↓ 22 1:46 PM • Mar 16, 2026 |
Tuesday 17th March 2026
Wednesday 18th March 2026
| 1 | OpenAI launches GPT-5.4 mini and nano, aimed at agents, coding, and multi-modal workflows, and offering near GPT-5.4-level performance at a much lower cost OpenAI Developers @OpenAIDevs We’re introducing GPT-5.4 mini and nano, our most capable small models yet. GPT-5.4 mini is more than 2x faster than GPT-5 mini. Optimized for coding, computer use, multimodal understanding, and subagents. For lighter-weight tasks, GPT-5.4 nano is our smallest and cheapest 6,486 5:09 PM • Mar 17, 2026 OpenRouter @OpenRouter The new @OpenAI GPT-5.4 mini and nano are available now on OpenRouter! In our early testing, mini’s increase in speed was useful for staying in the loop for coding agents, and results in better UX for chat apps that allow the models to perform agent tasks mid conversation. 278 6:15 PM • Mar 17, 2026 [GPT-5.4 mini and GPT-5.4 nano, which can describe 76,000 photos for $52 OpenAI today: Introducing GPT‑5.4 mini and nano. These models join GPT-5.4 which was released two weeks ago. OpenAI's self-reported benchmarks show the new 5.4-nano out-performing their previous GPT-5 mini model ... https://simonwillison.net |
| 2 | Gokul Rajaram(the godfather of AdSense) in his podcast with Harry Stebbings of 20VC fame breaks down the 8 defensible moats you need for your company to be successful in a world of AI. 1. Data (Proprietary and inaccessible) 2. Workflow (Deeply embedded operations) 3. Regulatory (Licenses and contracts) 4. Distribution (Exclusive proprietary channels) 5. Ecosystem (Third-party platform reliance) 6. Network (Marketplace liquidity density) 7. Physical (Infrastructure and atoms) 8. Scale (Low cost through volume) Play Video |
| 3 | Minimax-2.7 released The Shanghai-based MiniMax lab released MiniMax-M2.7 via API, excelling on benchmarks like 56.22% on SWE-Pro for software engineering and leading open-source models at 1,495 ELO on GDPval-AA. It handles a 204,800-token context, outputs at 60-100 tokens per second, and starts at $0.3 per million input tokens, making advanced agent tasks accessible. [MiniMax M2.7: Early Echoes of Self-Evolution In the months following the first release of our M2-series models, we received a large volume of feedback and suggestions from enthusiastic users and developers, which drove us to further accelerate the efficiency of our model iterations. With human productivity already fully unleashed, the natural next step was to initiate self-evolution of both the model and the organization. https://www.minimax.io |
| 4 | Anthropic rolled out Dispatch in its Claude Cowork desktop app, letting users pair their phone with a persistent AI session on their computer to handle jobs like file analysis or Slack searches while away. It starts as a research preview for Max subscribers, with Pro access coming soon, and runs securely in a sandbox after user approval. The feature rivals open-source tools like OpenClaw by offering easy, local automation without maintenance hassles or security risks, as noted by users like Ethan Mollick who say it covers most of what they needed. Felix Rieseberg @felixrieseberg We're shipping a new feature in Claude Cowork as a research preview that I'm excited about: Dispatch! One persistent conversation with Claude that runs on your computer. Message it from your phone. Come back to finished work. To try it out, download Claude Desktop, then pair 17.3K 8:35 PM • Mar 17, 2026 |
Thursday 19th March 2026
Astral to join OpenAI
OpenAI says it will buy the startup Astral, as the AI giant goes after the gorwing market for AI developer tools
We've reached an agreement to acquire Astral.
After we close, OpenAI plans for @astral_sh to join our Codex team, with a continued focus on building great tools and advancing the shared mission of making developers more productive.
openai.com OpenAI to acquire Astral \n\Accelerates Codex growth to power the next generation of Python developer tools](https://t.co/V0rDo0G8h9)
[Astral to join OpenAI \n\Astral has entered into an agreement to join OpenAI as part of the Codex team.\n\https://astral.sh\n\
Friday 20th March 2026
| 1 | Fal MCP Server is live Connect Claude, Cursor, or any AI assistant to 1,000+ generative AI models. Search models, generate images, create videos, check doc, create app : from a conversation. [Connect your AI to 1,000+ models with the fal MCP Server Today we're launching the fal MCP Server - a hosted endpoint that lets any AI assistant search, run, and chain 1,000+ generative AI models directly from a conversation. One command. No SDK. No docs to read. What is it? The Model Context Protocol (MCP) is an open standard that lets https://blog.fal.ai |
| 2 | Composer 2 is now live in Cursor, pairing frontier-level coding benchmarks with standout pricing: $0.50 per million input tokens and $2.50 per million output tokens. Cursor reports major jumps over prior versions across CursorBench (61.3), Terminal-Bench 2.0 (61.7), and SWE-bench Multilingual (73.7), making this a notably very very strong price-performance launch for coding AI Cursor @cursor_ai Composer 2 is now available in Cursor. 9,436 4:30 PM • Mar 19, 2026 [Introducing Composer 2 · Cursor Frontier-level coding with strong CursorBench results, higher token efficiency, and a faster default variant. https://cursor.com |
| 3 | via Packy McCormick of the NotBoring newsletter [260K+ subscribers[0m There is a tremendous amount of progress happening in World Models. Multiple labs have raised more than $1B. WMs were the star of GTC. They are a real path to embodied AI. So Pim de Witte [0mof General Intuition fame which raised a seed round of $133.7 million[0m & I wrote a comprehensive 19k word overview of World Models. [World Models: Computing the Uncomputable A Co-Written Essay with General Intuition's Pim DeWitte https://www.notboring.co |
| 4 | OpenCode Removes Claude Max Plugin After Anthropic Legal Requests OpenCode, a flexible open-source AI coding tool that works in terminals and IDEs, will drop its Claude Max plugin in the upcoming 1.3.0 release following legal requests from Anthropic. The company prohibits routing consumer subscriptions like Claude Max—priced around $200 monthly—through third-party apps, enforcing this since January 2026 with errors and suspensions. Users can still access Claude models via official API keys or platforms like Amazon Bedrock, while OpenCode supports over 75 providers including OpenAI and Google Gemini. dax @thdxr opencode 1.3.0 will no longer autoload the claude max plugin we did our best to convince anthropic to support developer choice but they sent lawyers it's your right to access services however you wish but it is also their right to block whoever they want we can't maintain an 8,354 8:33 PM • Mar 19, 2026 |
Saturday 21st March 2026
| 1 | via Robby Yung 𝙋𝙖𝙨𝙨𝙡𝙪𝙤 @passluo Ladies and 乡亲们, Let's welcome the world's first Claw running on Nintendo Switch The Childless Cat Lady Tutorial will arrive tomorrow 139 2:29 PM • Mar 20, 2026 |
| 2 | In a Claude blog post, Anthropic product leader Cat Wu details four key changes for PMs facing fast-evolving AI: short 'side quest' sprints, demos over documents, revisiting features after each model update, and embracing simplicity to avoid failures. Her Claude Code team turned quick prototypes—like desktop support and todo lists—into launches after positive internal feedback, while evals now measure model performance. External PMs from Google, Microsoft, and others echo the gains, noting specs are evolving into the product itself as roles blur in collaborative workflows. cat @_catwu The PM playbook was built on an assumption that the technology underneath your product is roughly stable With the current pace of model progress, this is no longer true. Here's how we've evolved the PM role: 55 10:14 PM • Mar 20, 2026 |
The cover image of this newsletter via generated via the Nano Banana 2 model within the Freepik tool via the following prompt
A serene coastal vista at noon, showcasing islands with imposing, densely green mountains against a teal sky. No clouds are present. The sun is a brilliant white, positioned directly overhead, creating a high-contrast scene.