merve
@mervenoyann ·
"I run my models locally" this is how it looks like in 2026
you get model recommendations and zero excuses not to run models locally
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 · 2978 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 .
Kieran Klaassen@kieranklaassen ·
I love sweep!
Ben Burtenshaw@ben_burtenshaw ·
You can win $20K in cash prizes by building a demo, training a model, or creating a dataset with Beam by @reflection_ai!
Prizes are split across 3 winners that we'll choose after the challenge ends on Monday, November 2nd.
Plus, we're giving $25 hub credits to the first 400 to take part, just find the link in the page below.
Rules: Build something cool with beam and drop it on the discussion! Anything that might be helpful to the world or a cool demo is allowed:
- You could make a demo that uses the Beam model on Spaces - You could quantize it for your favorite hardware - You could generate a synthetic dataset for your favorite problem - You could build an RL environment with or for Beam - You could benchmark Beam and share the results. - You could surprise us!
If you build something with Beam, then share it on the Hub to enter the challenge. We’ll collect all the links, select the best, and reach out to winners.
https://huggingface.co/reflection/Beam-1.0-501B-A23B/discussions/5
Wes Bos@wesbos ·
3d modeled his hacked photo frame. This guy is KEEPING UP
Derya Unutmaz, MD@DeryaTR_ ·
This is a remarkable explanation of the significance of OpenAI’s model in refuting the Exponential Time Hypothesis (ETH), something I knew almost nothing about until now!
One extraordinary side effect of AI making discoveries in highly specialized fields is that it also makes those discoveries accessible to the rest of us. We get to learn about ideas that were previously almost impossible to understand without years of specialized training.
Perhaps mathematicians should have spent more time explaining these fascinating concepts to the public instead of complaining about AI solving difficult mathematical problems! Yet even now, I see some of them posting in such impenetrable jargon that only a handful of specialists can understand what they are saying. It often comes across as intellectual gatekeeping, as though explaining complex ideas to ordinary people were somehow beneath their intellectual standing.
And I suspect this is another reason why AI makes some of them so uncomfortable. AI is increasingly capable of solving problems that challenge even the brightest experts, while also explaining those problems and their significance in ways that people with little background in the field can understand.
That is an extraordinary transformation. AI is simultaneously pushing the boundaries of human knowledge and breaking down the intellectual barriers that have historically kept much of that knowledge inaccessible.
The ability to understand and appreciate profound ideas should never be the exclusive privilege of those who have spent decades mastering a field. AI is beginning to make that possible for everyone.
Arena.ai@arena ·
Exciting news: MAI-Image-2.7-Preview from @MicrosoftAI just landed at #3 overall in the Image Edit Arena with 1470 pts! That puts Microsoft AI at #2 among labs.
This release is a +45 pt improvement from its previous variant, MAI-Image-2.6 at #7.
MAI-Image-2.7-Preview also improved across every category compared to 2.6: - #3 in Product, Branding & Commercial Design (up from #8) - #3 in Portraits (up from #9) - #3 in Photorealistic & Cinematic Imagery (up from #6) - #4 in Text Rendering (up from #11) - #4 in Cartoon, Anime & Fantasy (up from #9) - #5 in 3D Imaging & Modeling (up from #10)
Congrats to the @MicrosoftAI team on this release!
Mitchell Hashimoto@mitchellh ·
JetBrains IDEs are now powered by libghostty! This brings huge feature and performance improvements to JB terminals! Based on their own annual reports, that's over 12.5 million active users across all platforms about to get a libghostty-powered terminal. https://blog.jetbrains.com/platform/2026/10/jetbrains-terminal-now-uses-ghostty/?utm_campaign=jb-ides-263-release&utm_medium=social&utm_source=x
I'll repeat myself: libghostty is and was always the primary goal of the Ghostty project (I said this in the first public talk I ever gave about Ghostty). The importance and impact of a free, liberally licensed, high quality terminal emulator library that runs anywhere and everywhere dwarfs any single terminal emulator application on its own.
I've estimated Ghostty's (GUI, not libghostty) own users at around a couple million. Overnight, 6x that are using a libghostty-powered terminal. From one integration.
Happy to see this, happy to hear JetBrains had a positive experience, and looking forward to continuing to improve and ship a stable libghostty tag soon!
Riley Brown@rileybrown ·
Yes… Obviously! AI is an amplifier.
I think 300 people have DMd me about video editing with Opus. “What skill should I use?”
And I always tell them. Don’t “use” a skill “Learn a skill”… Take 20 hours over 2 weekends to learn premiere pro / davinci resolve or CapCut. Make a video manually. Learn the vocab. Learn pacing. Learn jumpcuts. Learn to make a good video.
Then you’ll be able to take the expressions you’ve learned and turn it into repeatable workflows.
- Riley Brown
@rileybrown ·
My dream AI agent tool:
- Opus 5.5 Level Model (never delegates to worse) - Spins up cloud and local code sessions - Dots level real time voice / tool use - Codex level Computer use on local sessions - Claude Artifacts quality decks, docs and design - Mobile app is UI of duet - Inline mini apps (What Grokbot is doing with email and slack, but do this with alot more tools) - Webhooks (Like Grokbot has) - Built in social media access across all platforms - Plugins (of course)
In summary: I want an always on, frontier model powered employee that can use its own computer, and can use my computer 24/7 that self organizes all of its tasks into channels so that its easy to keep track of all the things we've done together. And I want to be able to control this agent on mobile, web, desktop, and most importantly real time voice.
Sophia Yang@sophiamyang ·
$18K + a free @nvidia DGX Spark 🤩
Join our @FireworksAI_HQ x @Microsoft x @nvidia SF buildathon.
To win: Pick a task, build an eval & beat a frontier model on quality, latency or cost.
Sign up here: https://luma.com/fireworks-8kmc
- Wes Bos
@wesbos ·
3D Model Anything scanner app is now in the app store! Give it a shot an send any PRs on github https://apps.apple.com/ca/app/3d-model-anything/id6825843398
Artificial Analysis@ArtificialAnlys ·
Microsoft's MAI-Image-2.7-Preview takes #2 on AA-Image-Editing v2.1, close behind OpenAI's GPT Image 2.5 Sunburst (max) and a clear step up from MAI-Image-2.6
MAI-Image-2.7-Preview is Microsoft AI's newest image editing model, succeeding MAI-Image-2.6, released in August.
In our testing, MAI-Image-2.7-Preview scores higher than MAI-Image-2.6 on all 7 editing actions we measure, and ranks #1 on 4 of them: Scene & Style Edit, Text or Symbol Edits, Reasoning-Based Edit, and Enhancement & Restoration. This new model improves on MAI-Image-2.6 with the biggest gains in Composition & Framing (+57 Elo), Text or Symbol Edits (+49) and Scene & Style Edit (+48). We see its best use cases to be in Live-Action Film, Consumer and Frontier edits, where it ranks #1.
MAI-Image-2.7-Preview is available today in MAI Playground for image editing only, with text to image stated to be coming soon.
Congratulations to @MicrosoftAI on the release!
See below for our detailed analysis and example outputs of MAI-Image-2.7-Preview in the Artificial Analysis Image Arena 🧵
Henry de Zoete@HZoete ·
AISI's brilliant Chem-Bio team published two new studies on specialised biological AI tools yesterday.
One asks whether coding agents can rebuild a tool from just its paper; the other, whether filtering training data can keep a dual-use capability out of an open-weight model.
This is exactly why the team exists: to understand and evaluate the chem-bio risks of frontier AI and their mitigations, and to share those findings with governments and industry. Full thread 👇
Vox@Voxyz_ai ·
OpenAI released GPT-6.1 Terra yesterday. It's 𝟯𝘅 𝗳𝗮𝘀𝘁𝗲𝗿 𝘁𝗵𝗮𝗻 𝟲.𝟭 𝗦𝗼𝗹 𝗮𝘁 𝗵𝗮𝗹𝗳 𝘁𝗵𝗲 𝗽𝗿𝗶𝗰𝗲, only a few points behind on benchmarks, and the best fit for a Codex subagent.
Set it up like the tree: the main session runs Sol on high reasoning effort and only breaks down tasks and does the final check. On OpenAI's DeepSWE chart, 6.1 Sol scores highest on high.
Every subagent runs Terra. Reading code and looking up docs get low, editing code gets xhigh. In a third-party coding agent benchmark, those are Terra's two most cost-efficient effort levels.
fork_turns sets how many turns of the main session's conversation a subagent takes with it. Leave it unset or set it to "all" and the subagent inherits the main session's model and effort. It 𝘀𝘁𝗶𝗹𝗹 𝗿𝘂𝗻𝘀 𝗦𝗼𝗹 𝗼𝗻 𝗵𝗶𝗴𝗵, and you've set up the tree for nothing.
Send the tree and this prompt to Codex 👇
"Set up my Codex to match this tree: → Main session: default to `gpt-6.1-sol` on high reasoning effort. It breaks down tasks, delegates, integrates, and verifies. Handle simple tasks directly; don't delegate them. → Three project-level custom agents: explorer reads code and researcher looks up docs, both with model `gpt-6.1-terra` and model_reasoning_effort `low`. worker edits code and runs tests, with model `gpt-6.1-terra` and model_reasoning_effort `xhigh`. Reuse fitting ones if they exist, and create only the missing ones. → If worker fails the tests on the same problem two rounds in a row, it hands the problem back to Sol in the main session. → Put the delegation rule in AGENTS.md: when spawning a subagent, set `fork_turns` to `"none"` or pass only the last few turns. If `fork_turns` is omitted or set to `"all"`, the subagent inherits the main session's model and effort.
Before changing anything, check: → Roles that still use older models like `gpt-6-sol` or `gpt-5.6-sol`. List them. → Roles whose model is set only in AGENTS.md and not in their agent config: check recent subagent session logs to confirm which model each one actually ran.
Show me the changes first, and wait for my OK before writing anything."
- Sophia Yang
@sophiamyang ·
who is going to @NeurIPSConf this year? which one are you going to?
elvis@omarsar0 ·
"We saw a glimpse of the future of swarm coding."
Agent swarms will fundamentally change how we build software.
@axiommathai used a swarm of agents to formally verify the Classification of Finite Simple Groups, the largest proof in math history. That is roughly 15,000 pages by over 100 mathematicians, now about 50 million lines of Lean, verified in 24 days.
The agents claimed tasks, wrote proofs, reviewed each other, and merged their own work through GitHub. Along the way, they fixed hundreds of major errata in the published proof.
Kevin Weil 🇺🇸@kevinweil ·
Slowly, then all at once. A full Lean formalization of the classification of finite simple groups!
The original proof was 15,000 pages of painstaking work across 500 papers by 100 mathematicians. Now settled beyond a doubt.
https://axiommath.ai/research/the-enormous-theorem-formalized/
Dan Shipper@danshipper ·
video creation and game making are going to become widely distributed foundations of knowledge work in the same way doc and deck creation are today
Elon Musk@elonmusk ·
Starlink Brasil 🇧🇷
Santiago@svpino ·
A bunch of agents chose which tasks to work on, wrote proofs, audited each other's work, and merged their contributions.
This is what software development will look like in the future. Heck, it's pretty much what it looks like right now!
It's hard to comprehend the scale we'll reach when agents can work independently indefinitely.
Read this below.
swyx@swyx ·
at spacex ipo everyone was posting their cursor bag fumble.
kid stuff.
my turn.
i estimate this was a $500m bag fumble lol. and very probably wouldve been on claude code early (bc i've been a typescript cli maintainer). fml.
kache@yacineMTB ·
Calling a model open weights is like releasing a closed binary and calling it open binary
Thor 雷神 ⚡️@thorwebdev ·
⚠️ Hardware hacking is addictive! You will start buying tons of little gadgets and have great ideas and tons of fun!
Got my hands on this little @waveshare00 ESP32-S3-ePaper board and loaded it up with tons of fun features:
👂 3.5 Transcribe Live for smart realtime transcription 🧠 3.5 Flash-Lite for prompt rewriting and routing 💬 3.8 Flash for voice design & speech generation 🍌 Nano Banana 2 Lite for image generation 📻 Lyria 3 Clip for music generation
Combining @GoogleDeepMind models in this small form factor is such good fun! I'm the favorite uncle now 🙌
Peter Gostev@petergostev ·
We are really underestimating how efficient inference is at big labs. Anthropic's gross margin is 85%, so it costs them $0.60 / $3.00 to serve Opus 5.5.
This is less than what OpenAI is charging for 6.1 Terra and much less than Muse, Gemini, Kimi etc.
dax@thdxr ·
neocloud sells gpus to inference providers who sell to model routers who sell to ???
where is all this capacity for OSS models going to go without products that create demand
someone's gotta make the apps
Marc Lou@marclou ·
Mentions of “TrustMRR” on X and Reddit in the last 24h.
Not sure what’s going on 🤷
Alex Atallah@alexatallah ·
We've announced OpenRouter's first major APAC partnership: KT Telecom, Korea's largest telecom company.
Around 80 media outlets took part. KT will become a provider of sovereign Korean models on OpenRouter. Launch will be anchored by KT Mi:dm K 3.0 Pro, Upstage Solar and Motif, with plans to expand 10-15 models in 2027.
OpenRouter will power KT's Token Factory, which will be available to KT's enterprise clients.
- Dan Shipper
@danshipper ·
extremely good explanation
@levelsio@levelsio ·
This is what I mean
South Brazil is great
Then read up on the "Southern Cone"
It's why @peterthiel it's moving there
You have to see them all as separate from Brazil imho
nicksaraev@nicksaraev ·
"I tried really hard at a thing and failed, therefore it's impossible" is not a good way to live.