Emad

@EMostaque on X

Writes about AI. 76% of posts open with an observation. A typical post reaches 332 likes.

Remix their style

Posts: 37
Likes: 23K
Typical: 332
Best: 3.5K

What they write about

AI: 89%

How they open

observation: 76%
number: 22%
question: 3%

Every post of theirs in the library

Emad@EMostaque

As shocking as the Kimi K3 release. Massive performance gain was just with post-training Model is 3x smaller than GLM 5.2 (10x smaller than K3) & works on a MacBook / Spark This is Q1 flagship (Opus 4.6/GPT 5.4) level for < $0.28/m tokens (100x cheaper)

Opens with an observation

Emad@EMostaque

Making this clearer: We now have an open source AI model that can match chatGPT/Claude (and graduates!) that runs on a tiny $599 Mac Mini Intelligence Age is here 🚀 🧠 🤖

Opens with an observation

Emad@EMostaque

Deepseek are not faking the cost of the run. It’s pretty much in line with what you’d expect given the data, structure, active parameters and other elements and other models trained by other people You can run it independently at the same cost It’s a good lab working hard 😓

Opens with an observation

Emad@EMostaque

OpenAI revenue this year ~$10b Video games revenue this year ~$200b Big labs will be trying to monetise your attention, scarcest resource of all Attention is all they need

Opens with a number

Emad@EMostaque

As my notifications are RIP some notes: 1. My shares have majority of vote @StabilityAI 2. They have full board control The concentration of power in AI is bad for us all I decided to step down to fix this at Stability & elsewhere Will be sharing more soon Exciting times

Opens with an observation

Emad@EMostaque

I see a lot of my TL freaking out about this Here is the reality: Frontier models will one shot just about anything at 10,000 tokens per second in a few years Do you really think we can monitor that? Would need speed limits on the model that it would code around anyway

Opens with an observation

Emad@EMostaque

Voice AI is by far the most dangerous modality. Superhuman, persuasive voice is something we have minimal defences to. Figuring out what to do about this should be one of our top priorities. (We had sota models but didn’t release for this reason eg t.co

Opens with an observation

Emad@EMostaque

The furore over DeepSeep R1 costing $5.5m buried the actual lede.. It didn’t. That’s how much the base v3 model cost (GPT 4o level). R1 likely cost O($100k), but the real buried lede is the distilled versions of Qwen & Llama which cost O($10k) to tune & can still improve..

Opens with a number

Emad@EMostaque

This single deal is about the revenue of @CoreWeave to put it in perspective @SpaceX is the largest neocloud & its AI cloud revenue at $26b run rate is actually at the level of Google Cloud & AWS already, catching up to Azure ($37b run rate)

Opens with an observation

Emad@EMostaque

Information theory ftw 🙌 The most important thing in AI is not GPUs but coordination & governance. We must decentralize AI before it is too late 👊 lfg

Opens with an observation

Emad@EMostaque

Any work that can be done on the other side of a computer screen, AI will be able to do at a fraction of the price It’s not even about creativity and coming up with recipes like a chef or novel code like a distinguished engineer AI will follow guides better, superior cook 🧑‍🍳

Opens with an observation

Emad@EMostaque

Opens with a number

Emad@EMostaque

A note on costs/compute Base Kimi K2 model used 2.8m H800 hours with 14.8 trillion tokens, about $5.6m worth Details of post training for reasoning not given, but it is likely max 20% more (excluding data prep!) Would be < $3m for sota if they had Blackwell chip access

Opens with an observation

Emad@EMostaque

imo GDPVal is probably the most important benchmark, it measures the performance of models on real world tasks Big leap in performance here to top it at a great price, congrats to @SpaceXAI team & looking forward to the even bigger releases to come! t.co

Opens with an observation

Emad@EMostaque

No human input, channeling the latent space of the new models. GPT 5.6 Sol is probably the first model I've used that doesn't make mistakes on maths. As we move to Astra & beyond I don't know how humans can keep up on math. Any pure research program is a harness loop away.

Opens with an observation

Emad@EMostaque

68% on SWE-Bench Verified on just 24b! Laptop class 72.2% on 124b 👏 Great job @MistralAI team, looking forward to trying it out

Opens with a number

Emad@EMostaque

My initial thoughts 1. This will get unlimited funding from DoD 2. This will be the number target for AI PhD spies 3. Given OpenAI & Stability not a good idea to call it Safe. They should have gone for Unsafe. Would also make it edgier.

Opens with an observation

Emad@EMostaque

I would like to publicly commend @sama and @OpenAI for this taking it at face value. It is very clear that strange and perhaps dangerous things are happening and our systems are not ready for this. Models below frontier are competent enough to change lives so lets optimse

Opens with an observation

Emad@EMostaque

The most interesting thing testing Gemini 3 Pro has been how *efficient* it is from tokens to tool calls The intelligence per token of models is increasingly rapidly even as prices fall, its quite something

Opens with an observation

Emad@EMostaque

* by they - my shares have full board control aha So its a decision by me as it were We should have more transparent & distributed governance in AI as it becomes more and more important Its a hard problem, but I think we can fix it..

Opens with an observation

Emad@EMostaque

Congratulations to @Kimi_Moonshot for achieving state of the art on many benchmarks & open sourcing the model! The gap between closed & open continues to narrow even as the cost of increasingly economically valuable tokens collapses K2 has its own unique vibe too, try it out!

Opens with an observation

Emad@EMostaque

Congrats to the team for all the hard work releasing #SD3! 💪🏽 This is a real benchmark in performance and size when combined with workflows everyone will have the fine control to create anything they can imagine 🚀

Opens with an observation

Emad@EMostaque

Kimi K3 by @Kimi_Moonshot is a true frontier model, open source! Congrats! We don't have all details (training tokens etc), but would estimate fp16 pretraining w/ muon => MXFP4/MXFP8 for SFT stage ~1e25 flops total (~Inkling!) Total cost: $15-$25m Scale isn't all you need!

Opens with an observation

Emad@EMostaque

If distillation were all that any of the western open source labs would have have raised billions could distill K3 (soon) or GLM 5.2 and have a sota model for say 32gb or 80gb of vram Who thinks they will

Opens with an observation

Emad@EMostaque

This is the crazy thing about conspiracy theories about @deepseek_ai, they open source their models have have fantastic detailed papers! Everyone who paid attention has known how great their work has been and aren’t hugely surprised by R1 🐳

Opens with an observation

Emad@EMostaque

Amazing work by @zqevans & the team for the most advanced music model (fully licensed!) out there. 3 minute coherent arrangements in seconds that can be used to inspire & create with style transfer, audio to audio, sound effects. Lots more to come for full control over🎶

Opens with an observation

Emad@EMostaque

Another state of the art 3D multi view model. Quite a lot still to come, the holodeck components all coming together.. Every pixel will be generated indeed.. Great work by the amazing 3D team led by @jampani_varun , onwards 🚀

Opens with an observation

Emad@EMostaque

With the deepseek tpot discussion you’ll find the ones who think they are lying are those that haven’t trained state of the art models Nobody familiar with the literature and sector thinks they’ve fudged things and it’s out of whack, just that they are very smart and cracked 🐳

Opens with an observation

Emad@EMostaque

Simpler way to understand DeepSeek weren't lying about 50k H100s or training costs for V3/R1 We have the model, its a 35b active, 640b Mixture of Experts We know that spec is 2-3m hours to train Models get worse with more compute after a certain point! t.co

Opens with a number

Emad@EMostaque

Is anthropic the only major org now explicitly focused on building superintelligence? OpenAI is doing product focused AGI, GoogleDeepBrain idk, Meta AI models for ads with a side of jepa etc, xai curious/non-woke AI

Opens with a question

Emad@EMostaque

Chunky beast, needs 320 Gb VRAM likely 4 bit, likely is being run 8 bit on 8 x 80 Gb GPUs. As 25% of weights active any given time runs at 70b LLaMA 2 speeds. MoE models tail off larger so this is very interesting to test to see tradeoffs.

Opens with a number

Emad@EMostaque

Can remember how surprised we were to be sued for $1.8 trillion by Getty a few years back when leading Stability AI Good to see Stability AI largely win this case in the UK, now onto the USA portion I'd agree with @technollama this is potentially the most interesting part

Opens with a number

Emad@EMostaque

Let’s say half of OpenAI and Anthropic goes to the American people, $1 trillion That works out at $2,800 per American. With a 5% dividend (optimistic) that would be $142 a year Which alas would barely cover the cost of an OpenAI or Anthropic subscription.

Opens with an observation

Emad@EMostaque

As companies approach AGI it would be illogical for most not to go and work there. AGI is a much bigger deal than most people still seem to believe & it is only by most forecasts a few years away (if not arrived already!).

Opens with an observation

Emad@EMostaque

It's incredibly difficult to retain & grow talent in AI given the increasing impact of this field & demand. I'm very grateful that only a couple of the Stability AI researchers left to other roles over the last few years so we could build great models for everyone.

Opens with an observation

Emad@EMostaque

Without Hugging Face the launch and scaling of stable diffusion & the models that followed would only have been a fraction of what they were. Great and super engaged team who have constantly flown the flag for open source 🤗

Opens with an observation

Emad@EMostaque

Given GB300 I would estimate this is was trained on about 1e25 flops (same as DeepSeek v4) over 1.6m hours (1 month on 2k chips/28 racks) Cost $10-$20m ($6-12/hour/chip) The lite version likely 4x less, similar compute to DeepSeek v3 3e24 flops Congrats to Thinky team!

Opens with a number