Dylan Patel

Valentine's Day for people in ML I posted the most normie ones, but @simonguozirui thread has many bangers for those who are cultured My fave being Spec Dec, over fitting, k & l, MNIST ones https://t.co/GDkDfYKqvo

The best part about Le Stargate is that German power prices are gonna spike and grid stability is gonna tank France exports 101.3TWh a year to her neighbors, or enough for 11 Gigawatts of datacenters. Le Stargate ambitions are not quite that large... yet, but one day they may be

@Kaizou_Bhuvi_17 https://t.co/frahauJbH1

The anatomy of datacenters is like peering into the the heart of AGI or something Unsung heros of data center design are cooling folks The difference in architecture between various firms in the AI arms race is quite drastically different

@main_horse Illegal acts and gore disallowed is good.

The new OpenAI model specs allows for sexual content As we speak millions of third world annotators are being tasked with the kinkiest roleplays. Hundreds of thousands of ai judges are being spun up to provide RL for synthetic furries The market opportunities are endless https://t.co/ltDKyXUel4

The point of Elon's offer to buy OpenAI for $97.4B to try to fuck over the non-profit to for-profit conversion A standing offer from many accredited investors to buy the non-profit portion OAI prolly tanks the argument to IRS/states Which then maybe kills future fundraising

Took 3 Google's to get the score of the game meanwhile ChatGPT and Perplexity took a single try This is why Google is trash Almost as trash as Patrick Mahomes tbh https://t.co/B5fWjevbee Y'all dumb, all the results are for the Superbowl It should show the score when I say game score It knows the context of where I am, who I am, and what the game is, delivers links that have the info I want but without giving it directly to me

Mira: We're doing reasoning but product focused. Me: What's the product? Mira TFW https://t.co/DDUZeNpBNR

Everyone saying Schulman but the other is arguable just as if not more GOAT

You know your brain is broken when aerial images of rectangular buildings now get immediately inferenced to "oooh datacetner" and you start looking for cooling towers and power equipment

AMD reused our code + modified for the benchmarks that show they're faster than Nvidia for training Few major caveats They didn't use pytorch native attention backend They used out of tree attention for MI300, but didn't give NV the same benefit or even the main tree version 1/2 They coulda used main tree or FA3 and achieved way better perf They used an out of tree GQA attention, but didn't enable the pytorch native gqa for Nvdia. Nvidia likely retests with bunch of Megatron and Vboost profiles to show even higher perf 2/2 https://t.co/riIuq6X8Zn

Honored to be @lexfridman podcast with the homie @natolambert 5 hour podcast dropped and it's a banger covering a lot of different topics. We had so much fun riffing. The secret to the hostage images Lex takes are now revealed with this nato selfie https://t.co/nOcy20WPFQ

OpenAI people are hyping their drop so fucking much. OpenAI Deep Research is reasoning + tool use agent for deep research Benchmarks seem really good, but no comparison to Google's Gemini Deep Research. Looking forward to playing with it. https://t.co/t7YykUZirV Successful tool use with an LLM chugging for 10+ min is fucking wild. We really are gonna get agents.

@GlennLuk @SemiAnalysis_ And it's a fault on our processes for making table fast not checking. Not saying I blame him.

We did a lot of content regarding DeepSeek for clients this week but eventually had to do a newsletter post when my brother texted me about deep seal https://t.co/ecSdwrxaR7

@royale_cla19901 @curiousiter @ohlennart We never said distilled. We said reward model

So this is like 2T in loss market cap for a $6M training run (ignoring cost of research, ablations, distilled data from GPT, capex for their various clusters, etc). Imagine if China invests $300m in a training run. It would reduce the world's GDP to $0.

Deepseek V3 and R1 discourse boils down to this. Shifting the curve means you build more and scale more dummies https://t.co/gOfgpyHYR2

OpenAI should have been a religion not a non profit Imagine the tax savings Morman church in shambles

@Chris0xFFF @SenougaharA @kevinsxu The source is here. Many major firms buy our data. https://t.co/55856FJ6s6

Posting a factual analysis of what the regulations do including being regulatory capture by Microsoft against many competitors, hurtintg Oracle significantly, the unintended/intended negative externalities Chinese datacenters in Malaysia are what these rules are designed to stop

@CavetoCreation @jordanschneider @SemiAnalysis_ It's illegal generally. You didn't see the executive order?

We breakdown what the OpenAI Stargate Joint Venture is including the winners and lowers, and where the money actually comes from? We also have estimates for the breakdown of costs for the clusters and sizing.

Rick and Morty is for underachieving "gifted students" who are actually midwits and get high too much

Elon's jet is in Florida. Global Foundries jet is in Florida. Qualcomm's jet is in Florida. In case anyone was wondering what's going on with Intel... They are at Mar-a-Lago. Make America Great Again can only be true if Intel is saved.

Nvidia will be a robotics company at the end of the day, not just semiconductor. Few understand the moves they are doing at the lowest level from manufacturing up to software. Classic extend embrace extinguish. They are going to sell robots and this is from Jensen straight up

@lithos_graphein @fabknowledge So cute tbh https://t.co/Jh5KG4MKPR

Huge depth and breath from the team on covering the Earth shattering AI Diffusion regulations Huge impacts globally All the coverage on this has been shallow, but we quantified + dove into actual impacts Also easter egg for my twitter folks in the very last line of piece, IYKYK

Apparantly yesterday Jensen was kinda pissed and downed 3 scotches on stage at their event last night for AI in Healthcare lmfao Note this is hearsay, I wasn't there cause reading regs