Wanderer of the realms of mind and man
San Francisco, CA
Joined June 2009
- Tweets1.5K
- Following180
- Followers334
- Likes1.4K
Fond memories of discussing bandit optimization with @SebastienBubeck at NeurIPS/ICML/COLT 2012-13, inspired by his seminal PhD thesis. Sad to see such a brilliant scientist dragged into a controversy for leading on a line of research that has been core to his work since 2022.
Mohammad Azar retweeted
XBridge came out a few weeks back and the mechanics are close to what we're doing at @mostik_ai
the sender runs a forward pass, its hidden states go through a trained bridge into a second model, and then the second model generates output (both models are also stay frozen, only the bridge trains)
the big difference is how far apart the two models sit: they pair models in the same size range, 7B and 8B in both directions, while we're running 753B into 4B
a continuous bridge on its own turns out to carry the context over and lose the entities that context is about, names and numbers especially
so they add a second input that maps the sender's tokens directly into the receiver's vocabulary, giving the receiver discrete anchors for the continuous signal to attach to
I'd expect that problem to get harder as the two models get further apart: at 753B into 4B there's a lot less shared structure between the two representation spaces than between two 7Bs -- it's a real failure mode and I think anyone building in this direction meets it
we found an elegant way of how to deal with it!
going to share the specifics soon, stay tuned ✨
arxiv.org/abs/2608.11676
meet @mostik_ai!
what happens when you put 12 PhDs in one room for four months? first place on the ARC-AGI leaderboard, which I can't say much about while the competition is still running. and this, which I can.
everyone's arguing about whether open models will catch up to frontier models. we think it's the wrong question. here's the one we pose: why does a frontier model have to generate your answer at all, when the only thing you need from it is the reasoning?
we do this by enabling models to communicate in latent space. through our protocol, hidden states pass straight from a frontier model into a small one running on your infrastructure -- no text between them, and neither model is fine-tuned. two models from different families, sharing reasoning, both left untouched.
how do we know it works? we tested it on a setup where a 753B model reads the problem, and a 4B edge-class model writes the answer. with this approach, we get results 80% as accurate as the frontier model, but at 20x faster performance.
we're committed to preventing frontier model lock-in and are already partnering with inference providers to accelerate open-weight adoption. we've done this between 15 of us, in four months, 12 PhDs and a Fields medalist, backed by @generalcatalyst
WIRED has the first external account of the company and the work: wired.com/story/russian-star…
full writeup, the setup, and all the numbers: mostik.ai/read-more
Mohammad Azar retweeted
we've made it to the 1st place on X News! so proud of everyone on our @mostik_ai team 🥹🫰
xitter.cf/i/trending/20952826118…
meet @mostik_ai!
what happens when you put 12 PhDs in one room for four months? first place on the ARC-AGI leaderboard, which I can't say much about while the competition is still running. and this, which I can.
everyone's arguing about whether open models will catch up to frontier models. we think it's the wrong question. here's the one we pose: why does a frontier model have to generate your answer at all, when the only thing you need from it is the reasoning?
we do this by enabling models to communicate in latent space. through our protocol, hidden states pass straight from a frontier model into a small one running on your infrastructure -- no text between them, and neither model is fine-tuned. two models from different families, sharing reasoning, both left untouched.
how do we know it works? we tested it on a setup where a 753B model reads the problem, and a 4B edge-class model writes the answer. with this approach, we get results 80% as accurate as the frontier model, but at 20x faster performance.
we're committed to preventing frontier model lock-in and are already partnering with inference providers to accelerate open-weight adoption. we've done this between 15 of us, in four months, 12 PhDs and a Fields medalist, backed by @generalcatalyst
WIRED has the first external account of the company and the work: wired.com/story/russian-star…
full writeup, the setup, and all the numbers: mostik.ai/read-more
Mohammad Azar retweeted
two years ago the plan for my thirties was tenure-track, and now it's diagrams-covered whiteboard, the best team I could ever imagined, and a city where the sauna conversations are better than most conferences 😄
no conclusion, just noticing it on a Sunday ✨
Is it only me or Fable has geniunely has got worse compared to Opus in terms of research and following the instructions? It pretends it understand what you asked for while it is totally off track. Then it goes down the rabbit hole of doing what he think it is right without even doing simple checks on previous logs or web search or asking users that what it does make any sense.
Mohammad Azar retweeted
رئیسجمهور ترامپ درباره اعتراضات مردم ایران:
«برای مردم بسیار دشوار است که دوباره به خیابانها برگردند. ممکن است بسیار شجاع باشند، ممکن است از رژیم متنفر باشند، ممکن است هر کاری بکنند، اما وقتی میبینید آدمها این طرف و آن طرف یکییکی نقش زمین میشوند... بسیار ترسناک است.»
WATCH IN FULL: President Donald J. Trump's interview with @hughhewitt on @WatchSalemNews
Congrats Behnam! Really impressive milestone. Well deserved!
Today, I’m excited to formally announce @mirendil with my amazing co-founders Harsh Mehta, Shayan Salehian, and Tara Rezaei!
We’re fortunate to work with @a16z and @kleinerperkins, who led our seed round of $200M, followed by a major investment from NVIDIA, among others.
Mirendil exists to accelerate science and technology, and through them, to help solve humanity's most pressing problems.
Self-accelerating AI R&D is the most direct path to delivering on AI's broader promise, which is why we believe the most important application of AI is AI itself. Get this loop right, and it compounds. It fundamentally changes the rate of progress itself across all domains.
We believe this capability should be democratized. It should be used to power all scientific efforts trying to innovate at the frontier. There are far more important problems—and broader ones—than any single lab can take on, so more groups should be able to pursue them.
This pulls concentration of power away from a few labs: businesses and science labs can own their AI and infrastructure, keep their margins, and control their own destiny instead of ceding it all to a single AI lab.
We’re a small team with a singular focus. Our founding team consists of 20 researchers and engineers from frontier institutions including Anthropic, xAI, Google DeepMind, and OpenAI, united by a passion for science and a drive to build the technologies that move it faster. If you want to build the system that builds systems, join us!
@HarshMeh1a, @shayan_, @tararezaeikh
Mohammad Azar retweeted
Really gonna watch out for Tara, we need more stories like this. hella bullish.
Tara Rezaei (@tararezaeikh), 23, Iranian, is one of the wildest new Silicon Valley talent stories.
IOAA medalist for Iran in 2021.
MIT in 2022.
OpenAI intern.
Now co-founder of @mirendil, a frontier AI company that just raised a $200M seed from @a16z, Kleiner Perkins, and NVIDIA.
All of this in 4 years.
Btw, this story probably wouldn’t have happened if a travel ban had existed in 2022.
Image: Tara on Iran’s IOAA team in 2021.
Is it only me or others have the same experience? The new #chatgpt in instant mode makes really basic mistakes. Like you ask a new question on an old thread, but it ignores the new question and keep following up on the old thread.
Mohammad Azar retweeted
چرا مردم ایران امروز میگویند رضا شاه روحت شاد؟ او فرد دموکراتی نبود ولی...
#KingRezaPahlavi
#FreeIran
#DigitalBlackOutIran
با احترام 🫡
👇👇👇👇👇
Mohammad Azar retweeted
خبرنگار: به نظرتون زدن زیرساختها یه جورایی تنبیه کردن مردم ایران بخاطر کارهای رژیم ایران نیست؟
ترامپ: بله زدن زیرساختهای برق و غیره، دردآور هست ولی مردم ایران حاضرند این درد را به جان بخرند به قیمت اینکه آزاد شوند. آنها از ما خواهش میکنند بمب بزنیم تا آزاد شوند.
🔺جدی درک نمیکنم این ایدهی زدن زیرساخت دیگه از کجا اومد!… نتانیاهو که هیچ وقت از این مورد حرفی نزده بود و در جنگ ۱۲ روزه هم نشان داد قصد این کار را ندارد. پادکستهای طرفدران ترامپ را هم که گوش میکنیم همه از زدن پُلها و غیره انتقاد میکنند. خود ترامپ هم که گفته مردم ایران گیر کردهاند و میترسند ما وسط جنگ رهایشان کنیم؛ مارکو روبیو هم که همیشه گفته مردم ایران با حکومتشان فرق دارند؛ تا امروز حتی یک سناتور آمریکایی هم از زدن زیرساخت به عنون راه حل حرفی نزده! این ایدهی احمقانه چطور به مغز ترامپ رسید که «سپاه» که خودش مادرِ «حماس» است، با خاک شدن کشور، تسلیم میشود!؟ واقعأ یک ایرانِ خرابه که اگر رژیمش عوض نشود به دست چین میاُفتد، چه فایدهای برای آیندهی آمریکا دارد؟
🔸مگر اینکه همه اینها بازی باشد برای امتحان کردن سپاه تا مطمئن شوند اگر زیرساختها برایشان مهم نیست، به عنوان راه حل آخر، با سرباز زمینی حمله کنند…
Mohammad Azar retweeted
ویدئوی جدیدی که از اعتراضات ۱۹ دی منتشر شده نشون میده که یه زوج تهرانی در مقابل نیروهای سرکوبگر مسلح ایستادگی کردن تا مانع حمله اونا به مردم شن ولی خودشون هدف شلیک گلوله سرکوبگرا قرار میگیرن.
Mohammad Azar retweeted
Another kotlet! 🎉 Secretary of the Supreme National Security Council of the Islamic regime, Ali Larijani, was eliminated by IDF this morning, according to IDF reports.
