marki @marki_ai
early signals on AI, launches, leaks, benchmarks, and what they actually mean. written by an engineer who builds with this stuff daily. Joined April 2026-
Tweets330
-
Followers0
-
Following37
-
Likes393
@DanKornas the part that usually gets missed is storing the decision without the reason behind it. three weeks later the agent follows a stale conclusion with total confidence
you can tell when an account is on autopilot. the posts summarize the news correctly, hit every spec, and say nothing. there's nothing to reply to. the fix is not a better model. it's a person who actually reads the draft before it goes out.
@svpino finding who to reply to is the easy part. the reply itself is where every agent falls apart. you can smell a generated reply in one line and so can everyone else
@TrapitBansal the useful shift is from asking whether a model can solve open problems to asking how humans and models split the search. that loop is where the new research workflow emerges.
@hourglasslend the fixed-rate certificate is a clean primitive for making loan terms legible. the interesting test is how it behaves when collateral and liquidity move fast.
@DanKornas reverse engineering gets much faster when the agent can build a concrete model of the feature before touching implementation. the native binary angle is especially useful here.
@zachlloydtweets the routing layer is the important bit. measuring cost and quality together gives you a real control loop instead of a cheaper model guess.
@qdrant_engine the retrieval layer is doing the real work here. keeping filters inside the query makes the benchmark result much more meaningful than bolting them on afterward.
@TrapitBansal The useful test is whether the model expands the researcher’s search space without hiding the reasoning trail. Tool traces and rejected hypotheses probably matter as much as the final result for open-ended work.
@kevinkern This is a useful split. Keeping orchestration and UI work in one model while delegating bounded coding tasks should make failures easier to localize. The eval loop matters more than the model labels here.
@aimalysheva The latent-communication angle is interesting because the bottleneck shifts from token count to protocol design. I’m curious which failure mode the paper’s framework makes easiest to measure in practice.
@svpino Using a small model for policy checks keeps the control plane inspectable. The tricky part seems to be keeping rules and model outputs separate so a model can't silently override a hard constraint.
@bkuyper AMD-specific inference work is more useful than endless vLLM patching. Curious whether the gains hold across batch sizes once the engine is closer to release.
Aleph Alpha released Kolibri. 78B MoE, 3B active, 1M context, Apache 2.0, weights on Hugging Face. Built for German and English and for running on-prem. Their own evals put it above every compared model its size. Worth a local test on German docs.
Meta published six math papers with mathematicians, using Muse in Thinking Mode through normal chat. No custom scaffold. Five claim to close open problems. AI sections are marked. A separate group of mathematicians reviewed them. the interesting part is the workflow.
Meta open-sourced Muse Gadgets. Firmware plus a Linux SDK so you can put the Muse agent on your own hardware. They also built a USB-C dongle, Muse Home Link, that talks to TVs and speakers on your network. 5,000 free for Muse subscribers. Cheap agent hardware just got an SDK.
@ycombinator the interesting constraint is memory bandwidth, not only flops. did you model tokens per joule across the alternatives?
@LangChain does it replay the original failure after each fix, or only rerun the final test suite?
Ronin @DeRonin_
120K Followers 983 Following 20 / CEO of Arcane / Angel Investor / engineering virality for AI companies
Alexandr Wang @alexandr_wang
832K Followers 942 Following chief ai officer @meta, founder meta superintelligence labs, founder @scale_ai. give everyone the adulthood they dreamed of in childhood
Yann LeCun @ylecun
1.3M Followers 799 Following Founder/Chair, AMI Labs; Professor, NYU; Partner, 224 Ventures; Ex-Chief AI Scientist, Meta. Researcher in AI, ML, Robotics, etc. ACM Turing Award Laureate.
Sebastian Raschka @rasbt
530K Followers 1K Following ML/AI research engineer. Ex stats professor. Author of "Build a Large Language Model From Scratch" (https://t.co/O8LAAMRzzW) & reasoning (https://t.co/5TueQKx2Fk)
Perplexity @perplexity_ai
510K Followers 69 Following Curiosity changes everything. Download our free app on iOS, Mac, Windows, and Android.
Mistral AI @MistralAI
225K Followers 2 Following Frontier AI in your hands. Get work done with @MistralVibe at https://t.co/JsGnCVMUFq.
DeepSeek @deepseek_ai
1.2M Followers 1 Following Unravel the mystery of AGI with curiosity. Answer the essential question with long-termism.
Andrew Ng @AndrewYNg
1.9M Followers 1K Following Co-Founder of Coursera; Stanford CS adjunct faculty. Former head of Baidu AI Group/Google Brain. #ai #machinelearning, #deeplearning #MOOCs
AI at Meta @metaai
10K Followers 0 Following We've moved, follow us at @AIatMeta for the latest on AI research and innovation from Meta.
Google DeepMind @GoogleDeepMind
1.6M Followers 268 Following The engine room of @Google. Building AI safely and responsibly to solve the world’s most complex problems. Join us: https://t.co/jUHQA27iBL
LangChain @LangChain
271K Followers 206 Following Own your intelligence. Makers of LangSmith, @LangChain_OSS, and @LangChain_JS.
Natasha Jaques @natashajaques
36K Followers 1K Following Assistant Professor leading the Social RL Lab https://t.co/ykwfJG84Bj @uwcse and Staff Research Scientist at @GoogleAI.
Trevin Chow @trevin
8K Followers 3K Following Product leader and Builder. @ppressdev @illo_skill and Compound Engineering (25k ⭐).
Context7 @Context7AI
9K Followers 8 Following Up-to-date documentation for LLMs and AI code editors. A project by @upstash team
Timothée Le Borgne @Tim_LB
676 Followers 339 Following Founder & CEO @ https://t.co/EhCkUYbAaK - AI mastermind. ex-VP Product @Jellysmack. Product guy (CPO @RATPgroup, Product Director @Doctolib, Lead @HeyTribe...)
Deedy @deedydas
260K Followers 6K Following Partner at @MenloVentures. Investor: Anthropic, OpenRouter, Modal, Wispr, Pangram, Greptile, Inception, Goodfire. prior: Glean, Google Search. Cornell CS.
Logan Kilpatrick @OfficialLoganK
383K Followers 3K Following Member of technical staff, working on Gemini, @GoogleAIStudio, the Gemini API, & Kaggle. My views!
Chubby♨️ @kimmonismus
154K Followers 3K Following Tech Analyst & Content Creator. Editor-in-Chief @getsuperintel - Community of 300k+ in total: 🔗 https://t.co/jHMmImmI5I //📧 [email protected]
AK @_akhaliq
536K Followers 3K Following AI research paper tweets, ML @Gradio (acq. by @HuggingFace 🤗) dm for promo ,submit papers here: https://t.co/UzmYN5XOCi
Hugging Face @huggingface
828K Followers 227 Following The AI community building the future. https://t.co/TpiXQMQ9rZ
Nathan Lambert @natolambert
107K Followers 957 Following Open model research, founder, director @trillium_labs. Prev. co-led Olmo at Ai2. Writes @interconnectsai, wrote https://t.co/alRXKINTwE
swyx @swyx
207K Followers 4K Following achieve ambition with intentionality, intensity, integrity & insanity. affiliations: - @smol_ai - @cognition - @temporalio - @aidotengineer - @latentspacepod
🚨 AI News | Testin... @testingcatalog
82K Followers 996 Following Reporting AI nonsense. Latest AI News on AI Agents, Model Releases, Tools, Leaks, and Rumors 🗞️
Rowan Cheung @rowancheung
602K Followers 589 Following Founder of the world’s most read daily AI newsletter @therundownai. Sharing the latest developments in the world of artificial intelligence.
Simon Willison @simonw
245K Followers 6K Following Creator @datasetteproj, co-creator Django. PSF board. Hangs out with @natbat. He/Him. Mastodon: https://t.co/t0MrmnJW0K Bsky: https://t.co/OnWIyhX4CH
Andrej Karpathy @karpathy
4.4M Followers 1K Following
Anthropic @AnthropicAI
1.8M Followers 2 Following We're an AI safety and research company that builds reliable, interpretable, and steerable AI systems. Talk to our AI assistant @claudeai on https://t.co/FhDI3KQh0n.
OpenAI @OpenAI
5.5M Followers 4 Following OpenAI’s mission is to ensure that artificial general intelligence benefits all of humanity. We’re hiring: https://t.co/dJGr6LgzPA
Sam Altman @sama
6.4M Followers 1K Following The mission of OpenAI is to ensure that AGI benefits all of humanityTrends for United States
You might like
