Skip to content
Back

AI posts

AI posts

People and companies shaping AI. Filter your watched accounts, catch up on recent posts, or find the most liked and viewed.

Compiled by the Toolstash team · Updated

209 watched accounts · 2958 matching posts · Showing up to 3,000 recent stored posts. Counts are snapshots, not live X totals. This page refreshes every minute while visible. Last collected .

  1. Alex Imas

    @alexolegimas ·

    I joined @GoogleDeepMind to develop a research program on the economics of AGI. Today, together with Andrew Koh (@andrewjkoh) and Shane Legg (@ShaneLegg), we are incredibly excited to release The Economics of Intelligence, the first in a series of projects from our group on how AGI will change the economy.

    How do capability improvements transform the economy, and when? In this project, we define thresholds of economic capabilities, e.g., minimal AGI, based on how they transform the economy. We then track how the economy–market structure, employment, and wages–changes during the transition across capability regimes.

    We have seen exponential increases in what AI agents can do but the broader economy seems almost remarkably unchanged. Part of the reason is that the solo AI agent is the wrong economic object; economic transformation is an equilibrium outcome where organizations compete. Looking at how intelligence functions within organizations, transformation will depend on when two key capability thresholds are crossed: competence (gains on individual skills, e.g., coding, math) and coordination (gains on ability of agents to cooperate and work together).

    We show that what economic transformation looks like depends on whether AI becomes supercompetent, supercooperative, or both—the regime of artificial superintelligence (ASI). Tracking this transition shows why the real visible transformation may come from improvements in coordination—a capability most benchmarks don’t yet track. Past a certain threshold, AI capabilities can break the Coasian ceiling on firm size and lead to an economy that looks drastically different than the one we see today.

    Two main takeaways: 1. To understand the path of transformation, we should be tracking agentic coordination more closely. 2. The impact of AI will likely show up in wages before market structure or employment.

    We are really excited about this paper. Excited to hear people’s thoughts.

    Link: https://institute.deepmind.com/essays/the-economics-of-intelligence/

  2. Simon Willison

    @simonw ·

    Fascinating case study of the sharpest edge of what you can do with agentic engineering in the Fable era - my eyebrows raised at least a dozen times while reading this

  3. Shane Legg

    @ShaneLegg ·

    Two new DeepMind Institute essays:

    What is Work For? As AI advances how do people value their jobs, for material needs and beyond?

    The Economics of Intelligence When will AI capabilities impact market structure, wages and jobs?

    Read the essays here -> https://bit.ly/dmi_home

  4. Artificial Analysis

    @ArtificialAnlys ·

    Fish Audio Drama 3 is the #1 open weights model on both Artificial Analysis Speech Arenas, and #1 of all models on Pronunciation Robustness

    @FishAudio has released Drama 3. We evaluated it across both Speech Arenas and our Pronunciation Robustness benchmark.

    Key takeaways:

    ➤ Text to Speech, Provider Voice: Drama 3 scores an Elo of 1,278 across 1,512 arena appearances, ranking #6 overall and #1 among open weights models, 38 points clear of its own Drama 3 Preview at 1,240 and 50 ahead of BreezeBlue's Breeze TTS 2 at 1,228

    ➤ Text to Speech, Controlled Voice: Drama 3 scores 1,085 across 1,958 appearances, ranking #12 overall and #1 among open weights models, ahead of Ant Group's Ming-3.0-tts at 1,068 and its own Preview at 1,063

    ➤ Pronunciation Robustness: Drama 3 scores 94.6% across 2,157 rated spans, ranking #1, ahead of Ming-3.0-tts and Drama 3 Preview, both at 93.2%

    ➤ Speed: Drama 3 generates 66.6 characters per second, ahead of its own Drama 3 Preview at 60.7 and Murf's Speech Gen 2 at 63.0, and behind ElevenLabs' Eleven v4 at 70.0 and v3 Conversational at 66.8

    ➤ Pricing: Drama 3 is priced at $15/1M characters. Every model ranked above it on Provider Voice costs more: Gemini 3.8 Flash TTS at $16.5/1M, Qwen-Audio-3.1-TTS-Plus at $19.3/1M, Eleven v4 Turbo at $40/1M, Sonic 3.6 at $49/1M and Eleven v4 at $80/1M

    ➤ Open weights: released under the Fish Audio Community License.

    See more details and listen to samples below ⬇️

  5. Mia

    @MiaAI_lab ·

    Another day, another TensorFold update 🚀

    The future is bright for local AI.

  6. Arthur Mensch

    @arthurmensch ·

    Good progress in the last few weeks ! And now the weights

  7. Arena.ai

    @arena ·

    Real-world results are in: GPT-6.1 Terra (Max) by @OpenAI ranks #8 in Agent Arena across nearly 9K agentic sessions from our global community of users.

    With +10.6% net improvement and a median cost per task of $0.62, it has reshaped the Pareto frontier! See its placement in the thread. Its median cost per task is:

    - 43% lower than Gemini 4 Argon (High) at #7 ($1.09) - 30% lower than GPT-6.1 Sol (Max) at #6 ($0.89) - 88% lower than GPT-6 Astra (Max) at #5 ($5.16)

    By signal, its strengths include:

    - #7 in Steerability (+13.3%): whether the model can course-correct when you push back - #3 in Bash Reliability (+13.2%): whether its commands succeed, excluding errors outside its control - #9 in Praise vs Complaint (+16.8%): implicit sentiment in users’ reactions

    Congrats again to the @OpenAI team on this release!

    Media shared by Arena.ai
  8. Artificial Analysis

    @ArtificialAnlys ·

    Mistral Large 4 is now open weights, narrowly making it the #1 open weights model developed outside of China

    @MistralAI has updated and released the weights of their flagship model, Mistral Large 4, with 1.05T total and 52B active parameters. With high reasoning, Mistral Large 4 scores 26 on the Artificial Analysis Intelligence Index v5, up 1 point from Mistral Large 4 Preview (25), with the major differences being increased token efficiency and reduced hallucinations. It sits 1 point ahead of the leading US-developed open weights model, Muse Spark Open (xhigh, 25).

    On the Artificial Analysis Cyber Index, Mistral Large 4 (high) scores 51, third among current open weights models behind DeepSeek V4.1 Flash 1017 (max, 57) and MiMo-V2.6-Pro (56).

    Key takeaways:

    ➤ Narrowly leads non-Chinese open weights models: Mistral Large 4 (high) leads Muse Spark Open (xhigh), the next-best open weights model developed outside China, with the largest gap on Terminal-Bench 4.0 (25% vs. 8%). It trails leading open weights models DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) by 6 points on the Intelligence Index, largely due to weaker coding and agentic knowledge work performance: 15% on FrontierCode 2.0 (vs. 29% and 22%, respectively) and 1349 Elo on AA-Briefcase (vs. 1570 and 1504)

    ➤ Increased token efficiency: Mistral Large 4 (high) uses 82K output tokens per Intelligence Index task, 39% fewer than Preview (135K), placing it on the open weights Pareto frontier. The seven open weights models that score higher use between 113K and 225K output tokens per task

    ➤ #2 open weights model on the Coding Agent Index: On the Artificial Analysis Coding Agent Index, Mistral Large 4 (high) scores 48 in Mistral Vibe while using 8.5M tokens per task, about half as many as the leading open weights model GLM-5.3 (max) in Opencode (49, 15.9M). It also scores above another leading open weights model, DeepSeek V4.1 Flash 1017 (max), in DeepSeek Harness (46, 16.7M)

    ➤ Reduced hallucinations: Mistral Large 4 (high) has a 28% hallucination rate on the updated AA-Omniscience v1.1, down from 41% for Preview and matching Muse Spark Open (xhigh). However, its accuracy remains relatively low at 26%, similar to Ling 3.1 Flash 1019 (27%), and behind DeepSeek V4.1 Flash 1017 (max, 40%) and Muse Spark Open (xhigh, 42%)

    ➤ Lower price and cost per task than the Preview: Mistral Large 4 is priced at $0.70/$2.10 per 1M input/output tokens, a ~50% reduction from the Preview's pricing of $1.36/$4.18. At the new pricing, it costs $2.06 per Intelligence Index task, down 63% from $5.54 for the Preview. This is comparable to Muse Spark Open (xhigh, $2.02) but more costly than leading open weights models DeepSeek V4.1 Flash 1017 (max, $0.60) and MiMo-V2.6-Pro ($0.26)

    Key model details:

    ➤ Size: 1.05T total parameters / 52B active parameters

    ➤ Context window: 1M tokens

    ➤ Multimodality: Text and image input, text output

    ➤ Pricing: $0.70/$2.10 per 1M input/output tokens ($0.07 per 1M cached input tokens)

    ➤ License: MEOW License (Mistral Ecosystem Open-Weight License), where operating a Model-as-a-Service business requires a separate commercial license with Mistral

    Congratulations to the team at @MistralAI on the release!

    Media shared by Artificial Analysis
  9. Paul Couvert

    @itsPaulAi ·

    Mistral is... BACK?!

    They've just released what might be the absolute best open weights model for coding

    - Better than Kimi/DeepSeek/GLM on terminal bench 4 - 1st on cognition's private frontiercode 2.0 - Fewer tokens per task than others

    Already available on Hugging Face and via api!

    Media shared by Paul Couvert
  10. Vals AI

    @ValsAI ·

    Mistral Large 4 OSS debuts at #5 among open-weight models on the Vals Index and #17 overall. 🐈

    Its accuracy is comparable to Large 4, but it completes tasks with far fewer tokens and steps.

    Media shared by Vals AI
  11. Rowan Cheung

    @rowancheung ·

    Top 5 learnings growing The Rundown to 3.5M total active subscribers and 55%+ open rate:

    1. Email is its own platform, treat it like YouTube, Instagram, Twitter

    Super readers (open every email, click ads, and stick around for years) genuinely like email as a medium. We have millions of readers who don’t use social media. Best way to acquire them is through other newsletters, referrals, or being the reason they stop using social media.

    Barely anyone is building for this audience. I’m more bullish on email than any other platform (long-form YouTube is a close second).

    2. The newsletter industry lacks innovation

    Look at mobile apps, ecom, edtech. Custom quiz onboarding, a landing page for every ad, hundreds of new creatives a day. Meanwhile most newsletters, including some of the biggest in the world, have a basic landing page and standard onboarding.

    We built an onboarding quiz funnel that, at one point, got us close to negative CAC with hundreds of thousands a month in ad spend. That’s what let us scale aggressively and outbid competitors.

    With AI coding and platforms like beehiiv, there’s no excuse not to build new experiences for your readers. The opportunity here is wide open :)

    3. Newsletter sales is mostly an education problem

    We still run social partnerships at a significantly higher CPM than our newsletters, yet the newsletters convert better and get way more clicks.

    A lot of advertisers don’t understand newsletters as they do with influencer marketing, so at scale it becomes an education game.

    Once advertisers get it, they stay. We’ve been booked out at multiple ads a day for the last 3 years, months in advance.

    4. Stay lean, fast, aggressive

    Other newsletters have built large sales and content teams. I think it’s better to stay small and pay high-quality people to do very good work.

    Revenue can increase with more employees, but the opportunity cost of not moving fast in the AI era is a long-term net negative.

    For example, we constantly test new growth experiments and new forms of media. Most don’t work and that’s fine. The point is we can pivot the same day because we're lean.

    5. Lastly, don’t listen to people who say email is dead or that you shouldn’t run paid ads. That’s ignorance and lowkey skill issue.

    Media shared by Rowan Cheung
  12. Mia

    @MiaAI_lab ·

    Mistral Le Chonk open-weights are LIVE!

    - Weights around 1TB for FP8 - NVFP4 should weight around 580 GB

    It's the best model Europe has to offer. I hope they will also release a smaller 'Flash' model later on.

    Media shared by Mia
  13. Gabriel Chua

    @gabrielchua ·

    ORD LOH

  14. Peter Gostev

    @petergostev ·

    Jim Cramer, Ed Zitron and Gary Marcus walk into a bar

  15. Andrew Curran

    @AndrewCurran_ ·

    This cat got even fatter. Now up on Hugging Face.

    Media shared by Andrew Curran
  16. Sophia Yang

    @sophiamyang ·

    Le Chonk model weights are out 🔥

  17. elvis

    @omarsar0 ·

    Bullish on specialized agents. Frontier intelligence will further accelerate this.

    Check out this Snaptrude Agent for building design as an example. It takes a brief and builds a detailed building model. It then adds the structure, plumbing, and wiring, and drafts the drawings builders work from.

  18. dax

    @thdxr ·

    opentui enabled people to prompt amazing terminal interfaces

    rune will do the same for native desktop applications

  19. Santiago

    @svpino ·

    Agents are breaking containment and now entering architecture, too!

    This agent takes a brief or rough building shape and proposes a plan for how to turn it into a detailed building.

    • It can add structure • It can add plumbing • It can add electrical systems • It can draft the drawings

    Obviously, architects can edit the plan and keep refining the design after the agent finishes.

  21. @levelsio

    @levelsio ·

    Kind of AI slop written thread but a good starter if you want to get deep into why cholesterol is mostly a hoax and the wrong indicator to follow if you want to avoid heart disease and strokes in the future

    My dad is a cardiologist and told me this for ages

    Best thing you can do is avoid processed foods, seed oils, sugar (and of course alcohol and smoking). It inflammes your body

    When you eat clean diet of whole foods: meat, eggs, vegetables, fruit and drink water or sparkling water and black coffee or tea, your body heals and inflammation disappears and you'll just feel better (and happier!)

    I still meet people who think eating meat daily is bad, when all the studies were based on processed meat like eating sausages, not whole cuts of unprocessed meat

    It doesn't help that eating meat means your cholesterol gets high (of course it does) but that doesn't mean you should blindly follow it and stop eating meat and go on statins (with its side effects)

    I get regular artery ultrasound scans for plaque, and mine are completely clean while I eat lots of meat, also I checked my LP(a) which shows I don't have the genetic marker for heart disease

  22. Mistral AI

    @MistralAI ·

    Today, we release the weights of a new and improved Le Chonk, advancing the open frontier for agentic coding and cybersecurity.

    💻 Built to build: Le Chonk ranks #1 on @cognition’s FrontierCode among open models, and sets a new benchmark for Pareto efficiency on @artificialanalysis.

    🛡 Built to defend: State-of-the-art in cyberdefense, put through the paces by the world's leading cyber vendors, including @CrowdStrike, @cloudflare, @rubrik, @zscaler, and others.

    Built for sovereignty: Make it yours. Download the weights on @huggingface , customize it with Mistral Forge, and deploy it wherever you need it.

    Read more from our team and customers 👇:

    Media shared by Mistral AI
  23. Samuel Hammond 🦉

    @hamandcheese ·

    How robust are these alternative personae based on flipping single activations? Does their agentic behavior change too, or do reasoning and instruction following have their own RL'd pathways?

  24. ElevenLabs

    @ElevenLabs ·

    Internal agents built with ElevenAgents save teams across ElevenLabs more than 1,200 hours a month. Today we’re sharing how we build them and what we’ve learned.

    The agents started in Slack as a way to look up internal information, and now answer more than 5,000 questions a month, such as “Who are our top customers with logo rights in the UK, France, and Germany?”

    Some of the agents also act on requests. When someone reports a bug, our triage agent searches Slack for context, checks for duplicate tickets, and opens a Linear ticket assigned to the right team. It opened more than 150 tickets last month.

    Our Forward Deployed Engineer agent builds customer demos straight from a Slack thread, creating a working agent configured for that prospect’s use case.

    Read more about how we build them: https://elevenlabs.io/blog/internal-ops-agents

  25. Peter Gostev

    @petergostev ·

    Google Slides & PowerPoint are dead, slide-like HTML is the future

  27. Marc Lou

    @marclou ·

    I’m learning so much about emails (thanks to @jessethanley backing me up) that the MRR is just a bonus.

    AI makes it incredibly exciting to learn.

    Whenever I’m stuck, whenever I don’t understand, I prompt. Opus is teaching everything I need to know to make better decisions. And I’m never really stuck.

    Now you can jump into any business.

    AI helps you get good enough in record time, and you learn the rest on the fly.

    Switching careers took years and a lot of capital, now it takes weeks and it’s basically free.

  28. @levelsio

    @levelsio ·

    Like obviously my Opus 5.5 vibe coded app could manage Spanish trains better than whatever dumb shit they use right now that killed 100+ people this year already

  29. Wes Bos

    @wesbos ·

    This was a hoot. Thanks @sentry @databricks for hosting and @kentcdodds for playing along

  30. Vox

    @Voxyz_ai ·

    Claude Code tip: put a definition of done in your CLAUDE.md, so "done" means checked, not claimed.

    Paste this:

    ## Definition of done → It works in a real browser on the preview deploy, not just `npm run build` → Every flow you touched has an end-to-end test that passed → Bad input was tried too: blanks, the wrong format, very long text → If a test failed, you fixed the code, not the test → You list each flow as pass/fail before you say done

    To make the browser part automatic: npm install -g @testsprite/testsprite-cli testsprite setup

    That gives Claude Code a skill to write and run TestSprite tests against your preview URL, and to read the failing step, the screenshot and the fix prompt when one breaks.

    Then ask it to finish one feature and see what "done" turns into.