Sometimes I write something and then I realize I am not the most qualified to talk about it then I delete said thing and go back to work
Good habit to have when you're terminally online
AI & LLM
Ce qui est entré en base, sans jugement.
5674posts
88auteurs
706liens
1463 hde recul
104membres
Aucun recoupement sur 24 h pleines : ni un lien ni un tweet n'est relayé par
deux membres (5 liens, 18 tweets relayés). Le recul est
suffisant, donc cette fois le constat porte : ces comptes diffusent au lieu de
converger, et c'est la composition de la liste qu'il faut revoir — pas la patience
qu'il faut allonger.
Peter Yang a cité @petergyang
“Both of my parents got diagnosed with two cancers each within a week of each other.”
From @AshleyYuki, VP of Product at @OpenAI:
“I just decided to take a month [off work]. It was one of the best things I could have done in that moment to care for my parents.
You also have to
1) It’s really simple folks
2) find your one
3) get married
4) stay faithful
5) and don’t allow lust to conquer you
Using an agent that can change its models or code at any time without telling you is like building on quicksand
S
Aidan McLaughlin a relayé @shrimptron
i would give every single cent i have to live at the house in front of the mill valley public library. unfortunately i have like 2 thousand cents
Just a reminder that Opensource AI must win and that we haven’t won yet and must prepare for the upcoming planned rugpulls
Anecdotally, I think cheating is far more common in Asia.
Plenty of guys who have their wife and kids living in the US/Canada, while they do business in Asia and have another girlfriend or even family in a 二奶村.
Most of the time their wife knows about this too. Not saying
I just realized my RTX 5090 I bought for $3k is going for $6k now.
It’s impossible to build a new high end computer now.
Makes me wonder if I should snag a 7900 xtx as a backup before those go more crazy than they are.
Ahmad a cité @TheAhmadOsman
PREDICTION
1B model running in ODS beating Opus 5.5 in terms of intelligence and capabilities within 3 years
T
Mike Bradley a relayé @firstadopter
Nvidia CEO Jensen Huang's email to Nvidia employees:
To my colleagues at NVIDIA,
I owe my family and all of you an apology.
I was unfaithful to my wife. My actions hurt her and our children. There is no excuse for what I did. I am deeply sorry.
Throughout my life, my family
Peter Yang a cité @petergyang
“What does this mean for Dad? What question should I be asking right now?”
Here's how Ashley from @OpenAI uses ChatGPT during her dad’s appointments:
“As the doctor’s sharing things with me, a lot of the time I’m using it in real time.
They’ll mention a treatment or some word
Look at this nerd and how passionate he is about Local and Opensource AI, you probably should go to his talk on Tuesday at AIE Code
ALERT🚨🚨: AMD MI355 is 1.42x FASTER THAN GB300 NVL72. THIS IS ACHIEVED USING THE TILERT SOFTWARE, WHICH GETS 1330 p90 tok/s/user on long context workloads!
Thanks to AMD leadership for providing an MI350 cluster to TileRT maintainers, enabling them to achieve this great
D
Dimitris Papailiopoulos a cité @jackcai1206
Latent notes, let the model dream about summaries.
One.
Million.
AI Engineers.
thank you.
J
Dimitris Papailiopoulos a relayé @jackcai1206
After reading the MEMENTO paper (@vkontonis @DimitrisPapail) I wondered if compactions needs explicit text at all. Turns out latent compaction is possible!
AbridgeKV trains the model to compress its own output. Keeping 1/4 of the generated cache, it matches full-KV accuracy on
D
Dimitris Papailiopoulos a cité @DimitrisPapail
Pushes to public GitHub repos per quarter, 2020 to now. Doubled in a single year.
THE PRETRAINING TOKENS ARE FLOWING
D
Dimitris Papailiopoulos a cité @littmath
Look at that beautiful pretraining dataset growing.
Jeremy Howard a relayé @math_rachel
"When two things are different, you should not use one to mimic the other. It would've been a colossal mistake if Phil Jackson had pushed Jordan to mimic Pippen & Rodman.
"The same applies to AI & human skills. Using AI to expand what humans do is more productive" @DAcemogluMIT
Mike Bradley a relayé @swyx
Local AI is back on the menu!
incredibly hyped for @TheAhmadOsman's return to AIE CODE next week. follow him for some exclusive discount code drops
Peter Yang a cité @petergyang
“My family is the most important. For me, it’s a non-negotiable.”
From Ashley at @OpenAI:
“I’m lucky I’ve worked at employers that are understanding and supportive of these things. But I still have to set that boundary. I just block the time.
When I was younger, I always felt
P
J
Jim Fan a cité @DrJimFan
Real robot. Real time. Uncut shot. Autonomous.
Tomorrow.
R
Manny Bernabe a relayé @raymmar
Exploring a bit before web summit Lisbon. Who’s else is in town?
E
Eden Marco a relayé @elie2222
Just broke $1k MRR with @rakazobot!
Why?
It's the best open-source Grok Bot alternative.
Try it at. Excited for what's to come. We're just getting started!
clem 🤗 a relayé @Gradio
A Hackathon to kick-start your AI research career?
Håvard Ihle a relayé @bayeslord
We need to stop using generic terminology. AI is not a monolith. There are critical things we need to do with AI for economic prosperity, there are things we may have to do for reasons of national competition, there are things we will at some point need to do carefully for the
V
homanp a relayé @vortexau
I need to test more, but my experience so far matches this.
Never seen a refusal on Deepseek at all. Ever.
Håvard Ihle a relayé @bshlgrs
Toby’s analysis totally misses the point here IMO. He’s right that the graph of capabilities over time could be explained as a one-off jump in capabilities from the Mythos pretrain. But that graph is not the main reason why METR thinks AI R&D is now sped up 1.7x by AI. The graph
Ahmad a cité @TheAhmadOsman
Imagine a 1B model beating Opus 5.5
That's the future, prepare for it
Nous Research a relayé @Teknium
We surpassed 1000 community made plugins today!
Check them all out here:
I spent months of my life and thousands of dollars looking for the best AI coding setup.
I've finally found it - T3 Code.
So, I made a video showing my full $1,000/mo setup:
D
Dimitris Papailiopoulos a cité @PeterStone_TX
UT is a very special place, one with strong intellectual and academic aura and amazing faculty. If you’re on the market you should apply!
Trivia: got my PhD there 🐂🧡🤘
Nous Research a relayé @Teknium
We surpassed 1000 community made plugins today!
Check them all out here:
N
clem 🤗 a relayé @nvidiaomniverse
1 million downloads of SimReady assets on @HuggingFace. 🎉
SimReady assets help teams use 3D content across supported tools and simulation workflows, from robot workcells to AI factory digital twins.
Thank you to everyone experimenting, sharing feedback and building with
AI.Engineer 🔜 CODE a relayé @swyx
just arrived… best swag game in the business. u guys are in for a treat next week
Y
AI.Engineer 🔜 CODE a relayé @sunyuzheng
"I'm going to bed. Ship it when you're done."
What I told my AI agent nine days before @aiDotEngineer Shanghai. One agent and I ran the site, agenda and materials from Seattle: 44 releases, 322 posters and stage cards.
Thanks @swyx @SophiaXH25 @StandupPartners @Ninggao88
Tim Li a relayé @soumithchintala
Inkling 2 Small on OpenCode for free for a limited time, to help us improve our models further:
"I have a few sealed 5090s that I am thinking of selling to get a new car" is something I never thought I would tweet seriously but here we are lmaooo
D
Dimitris Papailiopoulos a cité @DimitrisPapail
I CLICKED NO
D
Of all the AI work that deserves to be called out as hype, weight-space programming should be near the bottom of the list. Yes some experiments had problems, but the idea is significant. It’s great that more people are thinking about it. Also what an incredibly clean RL env.
Dimitris Papailiopoulos a relayé @MillionInt
Misinterpreting bitter lesson and an incredibly short period when scaling transformers did most of the legwork has reward hacked brains of a whole generation of scientists.
The more compute you’re pouring in training you’re model, the *more* important it is what algorithm you’re
Funny enough intelligence is nothing but a software problem
Watch us not only reinvent the wheel but rediscover it
D
Dimitris Papailiopoulos a cité @DimitrisPapail
Weight-space programming 2023 edition: two grad students, a whiteboard, and a LoOpEd TrAnFfoRmEr that ran assembly.
Respect to @AngelikiGiannou and @shashank_r12 for riffing this hard on this before it was cool.
(yes, I’ve found another excuse to retweet this for the nth time)
i didn’t really understand meditations on moloch until i learned that fiji water is actually from fiji
J
Andrej Karpathy a relayé @loresjpg
Nice DeepSWE score! Does it survive a lowercase filename?
Introducing MaxxBench: 89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement.
Models clustered around 74% on DeepSWE end up 23 pp apart.
Good
LLMs as we know them are full of noisy data, trained with brute force computation on every layer, with the same mentality as throwing stuff at the wall and seeing what sticks
Now, take away the effecinencies and you'll see a clear path for Local AI to win
M
Oliver Cameron a relayé @MozarellaPesto
I don’t think this is unintuitive. I believe a lot in manifesting destiny. If all you can think about is how ai will be terrible, you will propagate that message into the zeitgeist and the models will distill it. If you can focus on the potential upside, you reduce p(doom)
After DeepSeek V4.1 Flash, Xiaomi MiMo is moving to a causal encoder-decoder for its next-gen models.
HySparse2: prompts run through only 25 of 49 layers. Decoder KV is projected from encoder states.
At 1M context vs MiMo-V2: 5x less prefill compute, 4.5x smaller KV cache.
Ahmad a relayé @swyx
Local AI is back on the menu!
incredibly hyped for @TheAhmadOsman's return to AIE CODE next week. follow him for some exclusive discount code drops
Amjad Masad a cité @balajis
Really enjoyed Kazakhstan — super AI pilled government. And a great place for the Network School.
S
leonie a relayé @SophiaXH25
Around 1,600 people joined us in @aiDotEngineer Shanghai over two days and I’m still processing that.
Seeing people share what they learned and who they connected with has been so rewarding.
To everyone who helped make this happen, and everyone who came: thank you. So proud to
J
CoRL Keynote. Hilton Austin. Monday Nov 9, 1:45 pm. 2 days countdown.
Something extraordinary.
Astra solved two audio/video bugs I've been battling for 18 months!!
Problem 1: dropped frames
> Occasionally in my videos, it would look like the video was dropping frame rate. Turns out, there's a setting in OBS called "buffering" which was set to auto. I thought auto
S
AI.Engineer 🔜 CODE a relayé @SophiaXH25
Around 1,600 people joined us in @aiDotEngineer Shanghai over two days and I’m still processing that.
Seeing people share what they learned and who they connected with has been so rewarding.
To everyone who helped make this happen, and everyone who came: thank you. So proud to
helping my son learn to play Zelda is an absolutely surreal experience...all the feels
Amjad Masad a relayé @ycombinator
Replit was building toward the AI coding era years before it had a name. In 2024, the company shipped an agent that could take a request in plain English, work through errors, provision its own infrastructure, and keep going until the software actually worked. The model companies
clem 🤗 a relayé @DynamicWebPaige
Such an awesome, empirically rigorous benchmark!! 🙌
Holding the model fixed and still seeing a ~2x swing in solve rate -- and ~11.7x difference in cost per solved task -- is proof of how much the harness actually matters. (And Harness Lens documenting prompts, tools, and
Strong pathways for local talent build a stronger NYC tech ecosystem.
Brendan Collins, Executive Director of the @NYCTechTalent, joins AI Dev NYC to share how the city is putting AI engineers in the classroom. The Pipeline connects employers, training providers, job seekers, and
HOLY SHIT Opus 5.5 made me this video
Entire thing built by Claude Code, from choosing quotes and clips to editing cuts, pacing, zooms, music, text overlays, and video & audio upscaling
Seriously watch this or bookmark it to for later
I
MiniMax Design (H3) a relayé @ImaStudio_ai
MiniMax H3 Fashion MV | 13 Cuts, 2 Min SCARLET RIOT 🔴
Made in Ima Studio with outfit changes, music, and editorial cuts.
Try it 👉
Full prompt below 👇
Swap in your own character and reuse it in one click.
Create a 15-second, 16:9, 24fps fashion MV based
V
Patrick Loeber a relayé @_mohansolo
Resetting everyone’s Gemini and Argon quota on Antigravity.
Lots of updates this week so want to make sure everyone can build for the weekend. Lots more to come next week!
Y
Håvard Ihle a relayé @YafahEdelman
Well now we know that they do indeed have something reasonably compelling, so it's more urgent than ever for them to share it.
T
Amjad Masad a relayé @amasad
There are so many Linux zero-day vulnerabilities these days, but on Replit we’re able to avoid 96% of them with our custom Kernel build.
Read about our new sandbox design.
Y
AI.Engineer 🔜 CODE a relayé @yoshi_flux
Robots won't learn to act from robot demos. They'll learn from video games.
Takeaways from @PimDeWitte's @aiDotEngineer talk on @gen_intuition (spun out of @Medal_TV):
- LLMs are turn-based. The real world isn't. GI's models take in frames and output actions continuously, in
C
Håvard Ihle a relayé @ChrisRMcGuire
The story keeps getting crazier. To recap everything that has been reported so far:
- The US intelligence community discovered that Jensen Huang, the CEO of the largest company in the world, was sleeping with Alice Huang, a Chinese spy and an executive at Megaspeed. (FT)
-
Ahmad a cité @TheAhmadOsman
"these people are about to know that I am running SoTA inference locally at home with my GPUs"
T
Håvard Ihle a relayé @tomekkorbak
I’m glad to see Congress asking OpenAI why we were fired as well as raising some of the safety concerns we’ve been raising. Grateful to @RepCasar and the other signatories for pursuing this. I’m very curious to hear the answer, too!
Tim Li a relayé @MTSlive
SITUATION EXPLAINED: Thinking Machines just shipped the top US open-weight model, and John Schulman says early versions tried to get around their guardrails.
• Inkling 2 Small scored 31 on the Artificial Analysis Intelligence Index, up from 13 for the original Inkling Small in
Wu Haoning a cité @ArtificialAnlys
actually impressive knowledge as such a small model
Wu Haoning a relayé @ArtificialAnlys
Hy4 achieves accuracy of 50% on AA-Omniscience v1.1, with a hallucination rate of 89%. On HLE-Diamond, it scores 33%.
MASSIVE NEWS
Local AI continues to spread like wildfire (because of GPUs)
Y
AI.Engineer 🔜 CODE a relayé @yizhou_yu
Best event I've attended recently - I made a lot of great connections and learned a lot.
Thanks @aiDotEngineer and @StandupPartners for organizing! Looking forward to attending again soon!
AI.Engineer 🔜 CODE a relayé @somi_ai
"Verify twice before answering" added 48% to cost per ticket. Zero change in accuracy. That's on Anthropic's own slide.
Lance Martin says "think step by step" and "be thorough" are debt now. If they're still in your system prompt, delete them.
Well, test it first. It's one
烟
AI.Engineer 🔜 CODE a relayé @teach_fireworks
@ryolu_ 在AIE 2026 Shanghai关于AI产品设计的几点分享非常棒，争取下次去现场参加，无奈这段太忙了，AIE @aiDotEngineer 是我非常愿意参加的线下线上高质量AI活动：
1. 让重要的假设看得见，
让一次修改小到可以放心试，
让交出去的工作，
在需要的时候也能拿回来。
2. 不是每一步都加确认按钮，
leonie a cité @aiDotEngineer
it was great to see how vibrant the chinese community is.
thank you to the organizers for putting together such a fantastic event!
Paging my favorite ChatGPT Images PM @adele__li
J
AI.Engineer 🔜 CODE a relayé @joshuacirjak
Covered the future of AI on VHS tape. Slide left to see the camera 📹 @daytonaio
D
AI.Engineer 🔜 CODE a relayé @dexhorthy
a year ago at @aidotengineer Code Summit I spoke about Research Plan Implement, how to apply context engineering to your coding agents, and why every concept in AI is doomed to suffer "semantic diffusion". That talk has over 600k views and changed the way engineers everywhere
Tim Li a relayé @teortaxesTex
Again, more unexpected capabilities
Thinking Machines is a very serious lab
T
Peter Yang a cité @jsnnsa
Alright, I just interviewed @jsnnsa and he blew my mind the most out of all my guests since probably @steipete.
Going to expedite shipping our episode.
Subscribe here to get it soon:
AI.Engineer 🔜 CODE a cité @swyx
衷心感谢, AIE 上海！
❤️👨💻🇨🇳
everyone congratulate @StandupPartners on hosting the first EVER AI Engineer in China!
People need to interact with LLMs as if they’re a representation of all humans and understand that their answers are a representation of the normal distribution of all humans (average)
That’s where anyone with an edge shines while using LLMs btw
K
AI.Engineer 🔜 CODE a relayé @KaranVaidya6
see you at the biggest ai event.
T
AI.Engineer 🔜 CODE a relayé @thorwebdev
Let's talk about talking to your coding agents. 🗯️
Excited to be back at @aiDotEngineer in SF next week together with @ivanleomk and @_philschmid for a workshop on building your own remote coding agents that you can talk to!
F
AI.Engineer 🔜 CODE a relayé @franciscojarceo
I'm hosting a workshop at the @aiDotEngineer Code Summit next Tuesday in SF! 🌁
Power your favorite harness with open inference using @vllm_project 's Agentic API.
We'll cover agentic inference, cloud agents, and agent swarms!
Come by and say hi! 🤠
Tim Li a relayé @teortaxesTex
wtf, this is excellent and actually somewhat scary. Thinky did this on a tiny compute and token budget, without any fancy architecture or RL tricks. This is the same level of achievement as 1017. Maybe better.
This is FINALLY what can be expected of top-tier ex-Frontier staff.
J
AI.Engineer 🔜 CODE a relayé @huozhi
很喜欢里面的一句话：「而是在每次一小小的选择里发现，原来我在乎的是这个」。🖤
工作角色的称谓可能只有那么几个，但如果我们无限放大我们自己选择的每个微小细节，就会发现其实我们都有特别的地方
I
AI.Engineer 🔜 CODE a relayé @ivanleomk
Excited to be speaking next week at @aiDotEngineer Code on how we've seen agents cheat on benchmarks at scale :)
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values.
Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of
T
Håvard Ihle a relayé @testingham
@lugaricano Hi Luis:
- concretely, the IRT is unique up to an affine transform, which allows us to make judgments of acceleration, though we cannot meaningfully talk about a level or proportional changes.
- the assumptions that get us this is that there exists a latent capability, & each of
Someone told me today that I say it like it is and that rubs a lot of people wrong
When did people become so fragile and with inflated egos that they cannot hear something opposing them?
Matthew Berman a cité @theo
Had my Grok @bot do an analysis on Theo's posts to see how he earned so much:
Feels like all the apps that can use models from different labs like @conductor_build, cursor and others are getting acquired by the labs themselves.
The ones remaining like @t3dotcodes and @orca_build are open source.
Wonder how this will all play out. Personally I think
V
clem 🤗 a relayé @varien
We've open-sourced Qwen Concept 27B, our first model.
It applies direct weight-space edits (inspired by the recent MIT CSAIL "Sparks of Weight-Space Programming" paper) to cut thinking tokens by 54% while improving GPQA Diamond performance by 3% versus the base Qwen model.
Dimitris Papailiopoulos a relayé @ml_angelopoulos
Scarcity breeds innovation. The relative lack of compute in China creates powerful commercial incentives to innovate on efficiency.
"For example, the Bytedance experiments suggest a variant of the Full Bandwidth Transformer might be comparable to a model of twice the size,
How does one of the largest open source projects in history do releases? 🦞
@dallinfinite walks through how OpenClaw ships: dev, beta and stable channels, a weekly beta train, and one agent session shepherding every release.
Tim Li a relayé @arcprize
Inkling 2 Small from @thinkymachines on ARC-AGI (Verified):
- ARC-AGI-2: 81.7%, $0.27/task
- ARC-AGI-1: 96.5%, $0.08/task
It outperforms every open-weight model we've evaluated on both ARC-AGI-1 & ARC-AGI-2.
Thinking Machines plans to release its full weights next week.
Strong pathways for local talent build a stronger NYC tech ecosystem.
Brendan Collins, Executive Director of the @NYCTechTalent, joins AI Dev NYC to share how the city is putting AI engineers in the classroom. The Pipeline brings together employers, training providers, job
Dimitris Papailiopoulos a relayé @_arohan_
I really love the full bandwidth transformer research - One of the beautiful works of art of this year!
“For example, the Bytedance experiments suggest a variant of the Full Bandwidth Transformer might be comparable to a model of twice the size, implying significant savings in
We partnered with @fmf_org and its member firms on a new framework to help identify when bypassing an AI system's safeguards could enable serious cyber harm, and prioritize stronger protections.
"...This framework is an important step toward a common methodology for evaluating
AI.Engineer 🔜 CODE a relayé @threepointone
I will be at AIE Code at the @AnthropicAI booth, buy a ticket, come say hi, I've got stickers and tokens
AI.Engineer 🔜 CODE a relayé @aiDotEngineer
The Bun rewrite story, from @jarredsumner of @AnthropicAI.
He's joining our Rewrites track at AIE CODE.
November 10 to 12, San Francisco.
Apply to attend:
Maxime Labonne a relayé @yb2698
Harness adapters training is now live!
We are training LoRA adapters on LFM2.5 / LFM3 from @liquidai and the recent Twil-9B model from @thewebAI using RL to make them more efficient and accurate on agentic tasks!
You can follow the training going live here:
Maxime Labonne a relayé @helloiamleonie
how we’re building decision models at @liquidai
(aka: “how we reverse-engineered Jev”)
谢谢 @aiDotEngineer shanghai for having me.
here is the thread version of my talk 🧵
AK a relayé @abidlabs
Where else should we integrate ML Intern?
N
Guillaume Lample @ NeurIPS 2024 a relayé @nex_ify
mistral’s student plan is insane right now
$6 for $165 worth of DeepSeek and GLM usage
you get:
> mistral pro at $5.99/mo (normally $14.99)
> $150 vibe code usage + $15 studio api credits
> access to deepseek v4.1 flash, glm 5.3, glm 5.3 flash and more
> full vibe coding
A new way to measure how AI agents handle the routine analyses of wet-lab science—BioBasicsBench.
Developed with Ligo, a biotech company focused on enzyme design, the benchmark draws on 100 tasks from real experiments to test whether agents can interpret raw lab data, identify
Great to catch up with @DavidCahn6 in Miami today!
Fun facts:
- David sent us a “blind term sheet” to get @coatuemgmt on the @huggingface cap table before we’d even spoken
- He then flew to Miami to convince us to make room in the round
- He kept backing us after moving to
D
Dimitris Papailiopoulos a cité @liorsela
This is exactly how Opus 5.5 sounds when I ask it to explain in plain words how the hell its new megakernel works.
Ahmad a cité @TheAhmadOsman
This might be my most controversial tweet since that one I wrote about Anthropic
LLMs are like shades of grey, they’re not black and white, and the sooner you realize this the sooner you’ll be able to predict the future
Biggest customer acquisition mistake:
Assumed customer profile: curious, loves you, grateful you built something for them, patient, rich, forgiving
Actual: furious, hates you, regrets they even clicked, stingy, miserable
How would you design your landing page now?
Manny Bernabe a cité @MannyBernabe
Be Cedric.
23. Marketing sophomore.
Scrolling TikTok.
Sees peptides and GLPs blowing up.
Only a couple of apps for it.
No technical background.
No Mac.
No clue about Xcode.
No problem.
Vibe codes Pep AI on Replit in under two weeks.
On a Chromebook.
Pushes it straight to the
AI.Engineer 🔜 CODE a cité @ZinkLu2
we have dubbed and translated @ryolu_'s AI Engineer Shanghai talk to English — with @ElevenLabs audio voice cloning, and VISUALLY with @OpenAI GPT-6 Astra (the slides translated and superimposed)
Our first ever dual-language video using YouTube's new
M
AI.Engineer 🔜 CODE a relayé @mihail_eric
Excited to be speaking at the @aiDotEngineer conference next week. I'll be discussing how to embed world-class GTM into vertical agents.
Special surprise for attendees (hint: I'm parachuting into the event from the Monaco plane)
G
Rowan Cheung a relayé @guyoseary
Rowan is the founder of The Rundown AI, the world's largest daily artificial intelligence newsletter with an audience of over 3.5 million people. We are honored to have him join the @sound_ventures_ family.
D
Dimitris Papailiopoulos a cité @JohnCLangford
John Langford writes about the end of AI Frontiers, the group I was a member of as well.
A
Rowan Cheung a relayé @alexeheath
Welcome Rowan, architect of the biggest AI newsletter in the world, to Sound Ventures!
We at Sound have a view that how you tell your story will only matter more and more as code becomes cheaper to create and capital becomes abundant. More to come...
F
Forward Future a cité @MatthewBerman
PDFs are cool again!
I
Ibragim a cité @reflection_ai
glad to see SWE-rebench V2 used in Beam’s training 🙌 more open-source coming soon
imo RL envs built on public data should stay open too!
Z
AI.Engineer 🔜 CODE a relayé @ZinkLu2
@aiDotEngineer 上海 2026 Day 2 的分享，质量真的太高了。录音我放评论区了，这里只谈谈我的感受。
上午这几场上来就棒极了，一开始是 @chlassner 分享的世界模型基础知识（原谅我英文真的不行，所以 Google Maps 是世界模型么，还是只是一个玩笑？），然后是 @liu8in 带来的炫酷演示。
K
AI.Engineer 🔜 CODE a relayé @kentcdodds
I'm going to be in San Francisco next week speaking at @aiDotEngineer about the three pillars you need for the software engineering flywheel.
It's time to develop trust and get real leverage out of agents.
Join my session!
G
AI.Engineer 🔜 CODE a relayé @Ninggao88
Without all our lovely and amazing volunteers and my team, we couldn’t make this happen! Mengying you are the best, and already look forward to the 2nd AIE China!
Where should we take next time? @aiDotEngineer @StandupPartners @swyx
F
AI.Engineer 🔜 CODE a relayé @facundozurdo
AI Engineer Code Summit will be simply amazing. Full of talks and workshops from really influential people working on the next generation of AI-powered projects.
If you are attending, let's meet! 🙌
If you don't have a ticket: What are you waiting for? 😄
Open-source is critical for robotics AI. It means a student, a researcher or a startup can build on years of progress and contribute something the whole community can use rather than concentrating everything in a few companies building a black-box robot.
That’s why we’re
easiest 7% gain ever
A
clem 🤗 a relayé @Gus_Crampette
We were at the @seeedstudio Shenzhen factory today with @antoinepirrone and @matth_lapeyre and seeing the Microduck production line run for the first time was something special.
We even had our own assembly stations 🙂
Nvidia Rubin Teardown! Direct from SemiAnalysis: in the lab, on the bench. Reach out to sales@semianalysis.com now to learn how you can get early access to the hardest hitting insights in DC/AI hardware.
J
I'm giving a Keynote talk at CoRL next Monday Nov 9, please come by at 1:30 pm at Hilton Austin!
We've been cooking 👀
N
Fei-Fei Li a relayé @neilsonks
We're in a computer graphics renaissance
Drawings, photos, renders are all something world models can "see into", world vertigo
Nous Research a cité @Meituan_LongCat
LongCat-2.5, the new multimodal model from @Meituan_LongCat, is free in Nous Portal for one week.
Try it free today at
clem 🤗 a relayé @joelniklaus
I burned a shitload of compute so you don't have to. Releasing the Harness Leaderboard.
Which harness should you run your open model in? I ran 28,849 trials to find out.
The leaderboard compares nine coding-agent harnesses on open models. In each comparison only the harness
S
Dimitris Papailiopoulos a relayé @shyamalanadkat
looped transformers are going to surprise in many ways
D
AK a relayé @HuggingPapers
Harness-S: Learning from Sibling Harnesses
A training method that uses complementary sibling harnesses to rescue failed agent rollouts and transfers those behaviors into a single policy that still works under the original harness.
This literally save me 10h a week.
N
AK a relayé @NetEaseYouDaoAI
Our Confucius4-R2T2（Real Real-Time Transcribe） paper is now on arXiv!
📄
R2T2 learns what to transcribe—and when to wait. 🎙️
clem 🤗 a relayé @LysandreJik
We’ve been hard at work trying to make models work with harnesses better. The best thing @yb2698 found: “harness adapters”. Small LoRAs trained with RL inside the actual agent harness they’ll be used with.
Same base model. Different adapter for pi, opencode, hermes, etc. Why?
clem 🤗 a relayé @natolambert
Looks like the smartest western open weight model by a margin. While much smaller than muse spark open or mistal large. While with a better license (apache). Love it!
D
AK a relayé @HuggingPapers
ByteDance Seed published a new dataset on Hugging Face: MLIP-liquid-corrections. It provides BSSE and ATM correction labels for 2M molecular configurations, derived from OMol25. This helps fine-tune ML interatomic potentials for accurate liquid simulations.
AK a relayé @EinsiaAI
Can coding agents replace traditional 3D vision pipelines and build a entire interactive worlds from code?
Introducing Einsia AI's Frontier3DBench — a unified benchmark evaluating coding agents across 3D Objects, Scenes, and Interactive Open Worlds.
- 145 tasks across 9
Fei-Fei Li a relayé @martin_casado
What the hell?!! I can't believe how far we've come. Atlas is so good. Look at this fucking splat!! Check out behind the DJ table. Totally consistent novel view. No ghosting. This is entirely generated from a single image.
I'm gonna turn this into such a sick VR experience.
D
João Moura a relayé @dosu_ai
"Every token that you get from them should be worth more to your company than what you're paying them." Listen to @joaomdmoura of @crewAIInc on getting the most out of every LLM request with @devstein64 🦻
R
Amjad Masad a relayé @Replit
Sumitomo Mitsui Finance and Leasing (SMFL) runs their business on Replit.
The Japanese-based company found that after using Replit, the development and testing period was reduced to one-third compared to the originally planned period.
homanp a cité @superagent_ai
While everyone else are busy battling out which abliterated model you should use, we've fixed the biggest problem agents hacking software: FALSE POSITIVES.
Superagent Red Team is reinforced from over half a million runs on 4532 targets.
cc @nikesharora
Håvard Ihle a cité @htihle
Kimi K3.1 Pro scores 9.9% on WeirdML, ahead of Kimi-K3 at 7.5%, but well behind Haiku 5.5, for example.
I ran it in Kimi-code with the fireworks provider.
I
Ibragim a cité @Alezander9
right now, you probably don't need orchestration.
you just need solo Opus 5.5.
I tested 7 different orchestration setups on 12 hard swe ml tasks
and all of them lost to a single Opus 5.5.
maybe 6.1 Terra for cheaper price
clem 🤗 a relayé @LeRobotHF
𝗟𝗲𝗥𝗼𝗯𝗼𝘁 𝘃𝟬.𝟳.𝟬 𝗶𝘀 𝗟𝗜𝗩𝗘 🚀🤖
From terabytes of data to prompting robots! Check out all what's new in the video below and join the 50,000+ developers & 300+ organizations in the LeRobot ecosystem 🤗
X
Ibragim a relayé @xdotli
Introducing FrontierPhysics - a benchmark evaluating agents for end-to-end physics research
Our core team from physics labs at Berkeley, Caltech, and Harvard recruited 42 physicists from 26 institutions. 28 of them published papers on Nature, Nature Physics, PRX, PRA*
1/7 🧵
R
AI.Engineer 🔜 CODE a relayé @ryolu_
gave a talk at @aiDotEngineer shanghai about about a question i keep coming back to: when we can make anything, do we still know what we want?
i want tools that help us find our own answers, not skip the process of forming them. not just more output, but more room to think,
Y
AI.Engineer 🔜 CODE a relayé @yizhou_yu
@liu8in of @HyperFrames_ gave an amazing talk at @aiDotEngineer today, and I enjoyed followup discussions with Bin, @WaterrrMiao and @wenbozhu0328
Bin's talk: HTML is all your agents need.
Quick highlights:
1. Exploit what the model is good at
The team used to ask agents to
S
AI.Engineer 🔜 CODE a relayé @SophiaXH25
We love Shanghai! ❤️❤️ Our attendees at @aiDotEngineer Shanghai are amazing 🥹
K
Genspark a cité @PedroBabreu
A script and a visual bible. That's all @PedroBabreu handed Media Studio, and CRUSH is what came back. If you make videos, this one is worth watching.
I tell people that Opus 5.5 is a small and specialized model
Try to give it a very difficult problem, it will fail
Out-of-distribution tasks are some of the things that only big models (think Astra/Sol) could do
R
Tim Li a relayé @0interestrates
Inkling 2 Small from @thinkymachines is now the best American open source model on DeepFinance bench
Inkling 2 Small scores higher than GLM 5.3, DeepSeek V4.1 Flash and Nvidia’s Nemotron 3 Ultra on long-horizon high finance workflows
Nous Research a cité @liamzebedee
That's hilarious but you should probably try Hermes which doesn't do that
BREAKING NEWS🚨: OpenAI just announced that it will deprecate GPT-5.6 Sol Ultra Fast. This is the same model that Cerebras publicly announced it ran at ultra-fast speeds for.
Setting up a Mac mini with Apple Messages that uses a local model to help the wife troubleshoot and add movies & shows to Plex was somehow among the top 3 thing I did in the past couple years lmao
A
Tim Li a relayé @ItsmeAjayKV
Inkling-2-Small now scores 31 in Artificial Analysis Intelligence Index.
That's one point below GLM-5.3 (max) 🔥
Knowing they are planning on full weight release soon is even more awesome.
Matthew Berman a cité @factory
“We will simulate a nation-state level attack”
🤯
Matthew Berman a cité @thinkymachines
Maybe this would be good for vibe editing?
Imagine wanting to ban the biggest technology unblock ever, which could enable curing disease, and help the US work our way out of our massive deficit.
Tim Li a relayé @johnschulman2
Proud to release our new model, Inkling 2 Small. Check out our blog post -- it includes a lot of detail about our training process.
To me, the most interesting part of this release was what we learned about alignment, which you can read about in the sections "Safety and
Can folks from @suno and @Tesla add the app to the cars by default. This shit is better than any radio.
Like it'll be amazing to just drive and voice dictate to create songs to listen to.
Folio Update v1.16.0!
Fun problem - I can't keep up with the PRs and issues being submitted 😂
Features:
> Adjustable pen widths
> OpenRouter support
> Multilingual scan recognition
> Back and Forward inside PDFs
> Print previews and advanced layouts
> Organized Settings and
leonie a relayé @ramin_m_h
Today @WSJ covered our work with @MercedesBenz. our flagship LFMs are now rolling out in vehicles in North America, running directly on the hardware inside the car to power in-car assistance! 🌊
this is one of the first deployments of local generative AI in the real world at
leonie a relayé @maximelabonne
Quantization-Aware Distillation (QAD) is a technique we developed that outperforms Q4_K_M in quality while using Q4_0.
Better quality, lower memory usage, and better compatibility for on-device deployment. Stay tuned!
Congrats to the SGLang team @sgl_project and @radixark on the amazing Day-0 Kimi K3.1 performance! Notably, the team shipped megakernel support out-of-the-box!
Tim Li a relayé @ar0cket1
they also seem to have done very large batch sizes for RL, almost the exact same size as mimo V2.6, around 3B tokens. But they did 36x more steps than mimo (at basically the same ish batch size).
this is probably the biggest RL run we have gotten some metrics on so far in OS.
D
Dimitris Papailiopoulos a cité @DimitrisPapail
yup
Tim Li a relayé @yacineMTB
Oh wait I'm reading their blog post . Finally. A solid mogging from America. Thank god
Tim Li a relayé @eliebakouch
inkling 2 small is extrememly impressive
it's a new pre train from scratch (17B active 594B total param), train fully in only a few dozen of days, with much much much less compute and data (only 12T token!) than any other model they are competing with. the base model ppl loss on
Tim Li a relayé @natolambert
Looks like the smartest western open weight model by a margin. While much smaller than muse spark open or mistal large. While with a better license (apache). Love it!
Tim Li a relayé @yacineMTB
I can't believe I tweeted a week ago saying that thinky machines was on to something. How did I know. Sometimes I predict things so well i still am surprised by the outcome
J
clem 🤗 a relayé @jeffboudier
We spent $2m so you don't have to.
And so you can learn how to train MoEs.
Mixture-of-Experts LLMs are what powers the AIs you use everyday, and few people in the world know how to train them at scale (a handful at each top AI lab).
This handbook distills everything we
R
Amjad Masad a relayé @ryantm
It was wild to see SaaStr using the Replit Agent as their Chief Marketing Officer. They showed us at our office how Replit gives them 3 new marketing ideas a day and is hooked up to all their systems so they can judge effectiveness. They say it's much more effective than the
Matthew Berman a cité @elder_plinius
I can’t stop thinking about this
so, how did everyone enjoy their bubble experience?
not bad right? humanity collectively threw a shit ton of money at something and opus 5.5 came out the other end.
very pretty result. would mortgage economy for again.
Tim Li a relayé @teortaxesTex
Best new base model?
We need more models with this being an explicit target.
Tim Li a relayé @lu__jasper
This the most interesting part of the release to me.
Inkling-Small was pretty difficult to post-train out the box compared to Chinese models like Qwen.
1/ Good to know it wasn't a skill issue on my part
2/ Looks like they specifically tried to address this issue in Inkling 2
You know how AI tells you something will take 2-3 days and then gets it done in 10 minutes?
When my video editors tell me it'll take 2-3 days I'm wondering how much they're using AI to help because I think this work can be done 10x faster now (and still with taste!).
Tweeting
Tim Li a relayé @teortaxesTex
> Pre-training done in just 68 hours.
Whoa
Amjad Masad a cité @ryanflorence
Replit is the only collaborative AI coding system out there.
Z
Amjad Masad a relayé @zarathustra5150
In the end, it was Tucker Carlson who was right.
R
T
S
Tim Li a relayé @Ber18791531
Proud to push long-horizon coding and autoresearch with the team!
A
AI.Engineer 🔜 CODE a relayé @adamm285
In 2009, @patrickdebois saw that software development and operations were disconnected and helped create DevOps.
At @aiDotEngineer, he called out the next crisis: "Context is the new code." We have 30 years of CI/CD, testing, and version control for code, but agent context is
Y
Tim Li a relayé @yiwenyuan98
I joined Thinky in August to work on coding post-training. Excited to see Inkling 2 Small launch today!
More exciting things to come!
F
Tim Li a relayé @fengyao1909
4-bit all the way from pre-training to post-training.
Pre-training done in just 68 hours.
Our model factory keeps shipping. Proud of our team!
Tim Li a relayé @thoughtfullab
Inkling 2 Small is a great American open model, showing frontier-level performance at autonomously improving other models.
With just 17B active parameters, it scores 31.87 on PostTrainBench, ahead of Kimi K3 (31.48) and GLM 5.2 (30.40), and on par with GPT-5.6 (Sol) and Opus 4.8
Tim Li a relayé @MiaAI_lab
Inkling 2 Small is an interesting model for me. On paper it's from the only US-based lab that actually looks capable vs GLM and DeepSeek.
594B total, 17B active.
And it's trained natively in NVFP4!
Open weights expected next week.
I'll 100% test it, & I might also do a recipe.
Amjad Masad a cité @Replit
There are so many Linux zero-day vulnerabilities these days, but on Replit we’re able to avoid 96% of them with our custom Kernel build.
Read about our new sandbox design.
Tim Li a relayé @teortaxesTex
OK pretty cool, a respectable "Flash" model
obviously crushes Muse Spark
Even has some "RSI" capabilities
Congrats Thinky!
Tim Li a relayé @grx_xce
This is the first-ever American open-weight model to hold the #1 spot on @DesignArena 🇺🇸
It's been remarkable to watch this team work with such intensity and care.
Huge congratulations, @thinkymachines, on a historic launch!
Y
Tim Li a relayé @yujia_bao
Spent the last few months working on midtraining. Really proud of what we shipped with Inkling 2 Small!
There’s still so much more to explore. If you want to push the frontier with us, we’re hiring!
Tim Li a relayé @miramurati
We share our approach to open-weights releases, how we assessed Inkling, why safety depends on both the model and the ecosystem it enters, and how testing, staged access, and stronger defenses can create a path toward greater openness.
H
Tim Li a relayé @x_hanyu
Thinky’s second model series is now live! 🚀
594B total parameters, 17B active parameters, and 9 days of RL training on 3T tokens.
More in
Tim Li a relayé @soumithchintala
the vision capabilities of Inkling 2 Small are a very nice upgrade from Inkling 1....just sayin
Tim Li a relayé @thoughtfullab
Inkling 2 Small can generate a full public equity diligence memo for just $0.69 using our finance harness.
We got early access to the model from @thinkymachines and have been evaluating it on DiligenceBench, where it’s performing remarkably well. It’s a 594B-parameter MoE model
R
Tim Li a relayé @ZhongRuiqi
It's a good model! try playing around with it and let us know if you have any feedback.
we are approaching frontier so your feedback will be super valuable for us to release a strong, use-able, and open model! 🤗
K
Tim Li a relayé @KasraMaza
Really proud of the team for this one! Lots of hard work went into this. More to come soon!
clem 🤗 a relayé @Thom_Wolf
18 of the 20 best open models today are Mixture-of-Experts. Dense is basically over.
And yet the know-how to train a good MoE still lives mostly in the heads of a handful of teams.
So we spent 800k GPU hours running the experiments ourselves and wrote everything down. Super
clem 🤗 a relayé @thinkymachines
Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.
Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.
clem 🤗 a relayé @soumithchintala
Our v2 small model is here.
I've been using it daily for a variety of things.
The pre-training ran for 68 hours on GB300s. Post-training was 17 million rollouts over 9 days.
Entirely pre-mid-post trained in NVFP4.
Available to train and fine-tune on Tinker right now.
Open
D
Dimitris Papailiopoulos a cité @DimitrisPapail
BTW this is the reason why i find it interesting when I hear claims "we did not distill". Everyone indirectly distills whether one likes it or not, unless you train on the common crawl of 2022 and use data vendors that don't use any AI tech.
J
Tim Li a relayé @MatternJustus
Congratulations to @thinkymachines!
I love to see the level of thoroughness and attention to detail when it comes to Inkling's evaluation methodology.
Also excited to see some of the auto-research evals we have been working on featured in the blog!
D
It seems like every day it becomes more within reach to pretrain a strong model (cf. many new labs sharing new models trained from scratch).
I suspect this is strongly correlated with the pool of high quality synthetic tokens on the web (GitHub, arXiv, etc.) growing and getting
Tim Li a relayé @DesignArena
BREAKING: Inkling 2 Small by @thinkymachines takes 4th place overall on DesignArena and is 1st place among open weights models with an Elo of 1376!
At $0.30/$1.00 per 1M input/output tokens, Inkling 2 Small is 15× cheaper on output than Kimi K3 by @Kimi_Moonshot. This model also
T
RT @cHHillee: Our infra has really improved quite a bit since the last time we released a model - this one took *barely* over 3 days to pre…
Tim Li a relayé @miramurati
Inkling 2 Small is available today on Tinker, with weights coming soon. It's a capable small model. We look forward to seeing what you make with it.
D
in the limit as all AI research is automated away and tasks get arbitrarily complex, the only job left for us is making sure reward hacking never happens. so yeah, safety researcher might soon be the last technical AI job left.
AK a relayé @Thom_Wolf
18 of the 20 best open models today are Mixture-of-Experts. Dense is basically over.
And yet the know-how to train a good MoE still lives mostly in the heads of a handful of teams.
So we spent 800k GPU hours running the experiments ourselves and wrote everything down. Super
T
Tim Li a cité @thinkymachines
Much more will come! Hello world, first of Inkling-2
L
Amjad Masad a relayé @lhchavez
how does @Replit ensure that all your data is secured at runtime and avoid _most_ Linux kernel vulnerabilities? join @jasondellaluce, @prydonius, @cgbrewster, and me for the first part of the story of how we went Linux Containers -> MicroVMs:
Find out how your agent breaks before your users do it for you.
Learn to design custom benchmarks with Mehdi Jamei & @andiPartovi (@veris_ai) at AI Dev NYC:
Build realistic test scenarios
Measure outcomes, not just responses
Keep evals repeatable
📅Nov 30–Dec 1, 2026 ·
I can’t think of many (any?) AI use cases where I can blindly throw tokens at a task and gain monetary value.
Mike Bradley a relayé @TheAhmadOsman
How I saw the future of Local AI
This was back in June, just 3 months after Qwen 3.5 27B
- Local AI = 90s of Linux
- The infrastructure not there (yet)
- We needed it OPTIMIZED for every piece of hardware
Now we are seeing a BOOM in Optimized Inference engines and Megakernels
F
Sam Altman a relayé @fouadmatin
Today’s GPT-6 Cyber release is a meaningful step for defenders: stronger vulnerability research and better alignment. We’ve already been using GPT-6 Cyber alongside GPT-6 Astra with Daybreak to harden our own systems, finding weaknesses, validating their impact, and developing
Maxime Labonne a relayé @ramin_m_h
Today @WSJ covered our work with @MercedesBenz. our flagship LFMs are now rolling out in vehicles in North America, running directly on the hardware inside the car to power in-car assistance! 🌊
this is one of the first deployments of local generative AI in the real world at
Fei-Fei Li a relayé @jcjohnss
We've released a demo of real-time world generation powered by our new Atlas model!
This is a spiritual successor to our RTFM demo from last year. It's a vision for the future of interactive media might look like, with frames streamed directly from a powerful generative model
D
Dimitris Papailiopoulos a cité @METR_Evals
The slope changed right around when Claude Code became usable for longer horizon research projects. Exactly when Opus 4.5 arrived.
Genspark’s growth in Japan took the stage in San Francisco.
Our Co-Founder & COO @sang_wen delivered a keynote at Japan Payment Innovation Meetup 2026 on Nov. 4.
Thank you to @smbc_group for bringing leaders in the Japanese market together 🤝
Nous Research a cité @0xdippo
Unfuck your hardware with Hermes
Matthew Berman a cité @advaith_sridhar
I love hard tech. @advaith_sridhar and team are using AI to discover new material science.
They just launched five Grand Challenges.
(I'm a small investor)
Matthew Berman a cité @claudeai
Rebuilding Intuit from first principles.
Don't ever think AI Labs won't compete with you.
Forward Future a relayé @MatthewBerman
Full breakdown of Weight Space Programming, the most insane AI research paper I've seen all year:
OpenClaw🦞 a cité @heyneighbor
So many agents
So many features
This thread is going to be 100 posts long soon
I guess I should be making more "(model) is INSANE" videos but screw chasing views all the time
AK a relayé @abidlabs
YOUR data deserves YOUR own, private model.
Now, let @HuggingFace's ML Intern build it for you.
Connect your data, files, mcps. And call it from Slack.
D
Dimitris Papailiopoulos a cité @skull8888888888
The bitter lesson is you have to do as much SFT as possible before you touch any RL. Unless you like spending money for no reason.
Artificial Analysis a cité @xdotli
Congratulations to @xdotli and the team at @benchflow_ai on launching FrontierPhysics! Beyond long-horizon agentic tasks in scientific domains, it’s great to see scoring that captures the nuances and judgment physicists care about
We're excited to partner with you to host this
AI is really wonderful because before only a handful of people could experience the pain and humiliation of building something without distribution.
Now millions of people can feel this!
With yesterday's elections results there is now no doubt that Local AI is the only way forward
clem 🤗 a relayé @NVIDIAAI
Great to see Nemotron 3.5 Transcribe trending #1 overall on @huggingface
Thank you to everyone for trying it, sharing it, and building with it. 💚
Go deeper into Intelligence Index v5 with the full presentation from last week's launch event in San Francisco
We introduced our biggest upgrade to the Index since v4, launched the new Progress Index, and shared what’s next for evaluation across 3D reasoning, personality, cyber
X
clem 🤗 a relayé @xenovacom
Real-time, local conversational AI is how we will democratize intelligence.
We just ported LFM3-Omni-2B by @LiquidAI to run fully offline on my iPhone 18 Pro Max at ~80 tok/s, powered by custom WebGPU kernels in a React Native app.
Should we release this on the App Store? 👀
View from Miami 🙌❤️🔥
clem 🤗 a relayé @HannesVonEssen
Microdock v2: As suggested by @matth_lapeyre, the robot now sits while docked, allowing all servos to be turned completely off.
D
AK a relayé @HuggingPapers
Turns Matter for Streaming Video RL
TurnWise is a new turn-level GRPO framework that gives each decision in a streaming conversation its own credit, improving when to answer vs. keep watching. It outperforms prior RL methods on three streaming benchmarks.
A
AK a relayé @AlexAtServal
Starting today, every engineer who joins Serval begins with a weeklong, $10,000 paid vacation. We’re building the next generational company. We need generational talent to see it through.
Our mission has always been to empower people across every team to turn their understanding
R
Manny Bernabe a relayé @Replit
How Work is Changing in 2027 (Replit for Business)
re: stock market and all the noise
Pessimists sound smart. Optimists make money
“Nat Friedman”
P
OpenClaw🦞 a relayé @particle_api
Our token efficient model particle-1 is now in OpenClaw
Google DeepMind a relayé @GoogleAIStudio
introducing gemini omni 1.2 flash, now available in ai studio and the gemini api
what's new:
- fast & 30% lower cost
- longer clips
- edit in conversation
- more natural speech
- more controllable shots
Maxime Labonne a cité @huggingface
Quantization-Aware Distillation (QAD) is a technique we developed that outperforms Q4_K_M in quality while using Q4_0.
Better quality, lower memory usage, and better compatibility for on-device deployment. Stay tuned!
🔴 We're Live! Replit for Business! Get in here!
N
Victoria Slocum a relayé @netbird
CrowdSec AppSec is a web app firewall that inspects every HTTP request for anything malicious.
As soon as you spin up a fresh server, you’ll have bots hammering /.𝚎𝚗𝚟 and trying every SSH password they know.
This is where CrowdSec comes in, blocking malicious IP addresses
Peter Yang a cité @petergyang
I have a lot of respect for doctors and have many friends who are physicians and nurses.
But it can sometimes feel like working with a coworker who doesn't share their email or phone number and only communicates through voicemails and faxes.
This is probably the most
Note to self
Deeply concerning to me... Delaware proposal to let agents run businesees... we should be very cautious here
"A Delaware AIC will be a separate legal entity whose day-to-day affairs are managed by an AI agent rather than a person. At the agent’s direction, the AIC can sue and be
Mustafa Suleyman a cité @Will4Planet
Been playing with this a bunch. Its super cool! Congrats @Will4Planet and the Planet team!
B
AI.Engineer 🔜 CODE a relayé @brianevanmiller
what firepower, it's going to be a banger
AI.Engineer 🔜 CODE a relayé @Xianbao_QIAN
AIE Shanghai is amazing
W
AI.Engineer 🔜 CODE a relayé @wey_gu
今天在 @aiDotEngineer 一整天都在说话，口水被榨干，收获超级多
还有好多朋友没见上
明天继续
我的 talk 会在 17:00 左右、xuanwo 哥哥 session 的后边
期待！
B
AI.Engineer 🔜 CODE a relayé @Braxxxx_Li
Was lucky enough to catch @ryolu_'s talk at @aiDotEngineer today, and got to chat with him a bit afterward.
What made me really happy is that a lot of scattered thoughts I've had for a while turned out to be very close to what Ryo was saying.
The one I agree with most: finding
N
AI.Engineer 🔜 CODE a relayé @nelsonpklark
@aiDotEngineer great first day in #shanghai #ai , interesting and vibrant Chinese community !
G
AI.Engineer 🔜 CODE a relayé @Ninggao88
@swyx you must wonder how you looked like in front of hull house of AI engineers in China! Great opening from Mengying @StandupPartners, golden words from @ryolu_ and such an amazing gaming demos from @MinakoOikawa Day 1 @aiDotEngineer
O
AI.Engineer 🔜 CODE a relayé @OpenProseVM
Compliance work is governed by language. So is the law it answers to.
At @aiDotEngineer NYC we showed agents doing that work: a quant model’s documentation goes in, audit-ready model risk documentation comes out, and every run is checked against a plain-English contract.
O
AI.Engineer 🔜 CODE a relayé @OpenProseVM
@aiDotEngineer Watch the full workshop:
Have a workflow like this inside your team? We’d love to talk.
Z
AI.Engineer 🔜 CODE a relayé @ZinkLu2
So here it is — @ryolu_ 's talk at @aiDotEngineer 2026 Shanghai.
I shot it handheld on a Xiaomi 15 Ultra, so there's some shake; I fixed as much of it as I could in CapCut. Handheld footage just isn't going to be sharp or steady.
Still, it's the best talk I saw on day 1 of AIE
S
AI.Engineer 🔜 CODE a relayé @Pinkelephantt
two things that stayed with me from @ryolu_’s talk at @aiDotEngineer:
i. when building becomes cheap, knowing what’s worth building matters more. keeping up with everyone else isn’t much of a reason to make something.
ii. not all friction is wasted effort. the false starts, the
AI alpha stops being alpha within 3 models generations at most (because they bake them directly into the models)
This is usually something I have to remind myself of whenever I feel the itch to “perfect” my workflows
I
Ibragim a cité @ibragim_bad
🚀 SWE-rebench V2.1 RL is out: 13,337 free RL environments for coding agents (20 languages, Harbor format)!
🔧 Fixed tasks, envs, and graders → fast, reliable rollouts.
Doing RL post-training or research? DM or reply. We can share data and RL envs early for free, so we can