i joined prime 12 months ago crazy what you can build in a year with a team like this `prime lab`

deepseek is passing the chatgpt detector tests with flying colors

it's interesting that 1.5 billion parameters is all you need to crush math competitions, but you need like 15 trillion to make the model be funny maybe humor is the right measure of true intelligence

grinding til 3am isn’t really a flex if you’re waking up at noon every day

TAke a look, y'all: IMG_4346.jpeg

idk about that, most of my friends know what it is

"don't write or run any code, just explain possible approaches and their tradeoffs"

i have evidence that openai is training on deepseek's user interface

it’s kind of funny that the “main ai guy” is basically nontechnical

my model picker still looks like this lol

i booked some flights today and i really don't know what part of the process could've been made much easier by ai. you just type in cities/dates, see options for times + prices, choose, enter your credit card info, then click buy. which part of that should be automated?

drop the “Deep”. just “Research”. it’s cleaner

rent is dropping significantly in all of the top 10 cities as ranked by magnitude of rent decrease

which is larger, 52.8 or 69.1?

Deep Research goes so hard if you spend 20 minutes writing your prompt

Sankey diagrams are a really intuitive way to visualize this kind of information flow. it’s a shame what happened to them

which AI division?

every dice roll i’ve ever done has either missed or beaten the expected value

you can literally make 6 figures for just having a hobby

if zuck is throwing around these kinds of numbers why doesn't he just buy anthropic lol

i’ve been significantly more productive since the whole ai thing really took off but it’s been like 10% using ai for stuff and 90% the whole “you have 2 years to escape the permanent underclass” thing

happy saturday for the price of a cup of coffee, you can rent a 4090 for the day and do some random experiments on small models

probably the highest-impact 6 lines of code i've written in my life

LLMs know way too many facts. you can ask Llama 1B about the history of pizza and it gives a decent answer. it shouldn't be able to do that. it should just do a google search. use those weights for something else

it’s kinda hilarious that the quora ceo has been on the board of openai since 2018 and they still missed this hard

this is how you know someone is cracked. no proofreading whatsoever, resume speaks for itself

anthropic employees should use twitter slightly more, openai employees should use twitter slightly less, xai employees should use twitter slightly differently

xAI Head of Product announces that Grok 4 is “the Antichrist”

imagine if gas stations didn't tell you how many gallons you were getting because car mileage was a trade secret and the gas station owned the car companies and you could either buy way overpriced gas per-mile or a monthly "max gas subscription" that turns off randomly sometimes

"IMO Medalist, Math PhD, or Masters in Data Science"

now this is a hiring page

I just read this new paper from Google and I’m absolutely buzzing 🤯 The core idea is almost offensively simple: ditch recurrence and convolutions, and use only attention. That’s it. And somehow…it unlocks a whole new regime of performance, scale, and simplicity. Here’s what blew my mind: - No recurrence, full parallelism. Tokens don’t have to march one step at a time anymore. Training lights up the whole sequence at once. Throughput goes way up, iteration cycles shrink. - Multi-head attention = multiple viewpoints. The model learns to focus on different relationships simultaneously. Syntax, semantics, long-range dependencies—captured in parallel. - Positional encodings without the baggage. You still get order awareness, but with zero recurrence overhead. - Encoder–decoder stacks that actually scale. Deep, clean, modular blocks with residual connections and layer norm that just…train. Reliably. - Results that speak for themselves. Stronger quality on translation benchmarks with dramatically better efficiency—and a simpler pipeline. Why this matters (right now): - Speed → strategy. When training is parallel and stable, you iterate faster, test more hypotheses, and ship better models sooner. - Quality → product. Long-range reasoning and richer representations turn into real-world wins: better search, smarter assistants, more robust generative systems. - Simplicity → leverage. Fewer moving parts, clearer abstractions, and a backbone that generalizes across tasks. This is an architectural blueprint, not a one-off trick. What I’m changing this week: - Refactoring any sequence stack I touch toward a Transformer backbone. - Re-thinking compute budgets around parallelism (bigger effective context, larger batches, faster turnaround). - Making attention the first-class citizen in modeling discussions—design defaults, not an afterthought. This paper feels like an inflection point. If you’re building anything with sequences—language, code, planning, you name it—read it, internalize it, and rethink your roadmap. The title isn’t marketing. Attention really is all you need. #AI #MachineLearning #NLP #Transformers #DeepLearning #GoogleAI #Attention #Research #ProductEngineering #Builders

the new llama 4 models are so advanced that they require versions of hf transformers that haven't even been invented yet

remember when yann lecun said this and everyone was like "shut up dummy" and scared him off this website

we’re truly about to enter a golden era of jane street and robinhood rugging the spreads on everyday americans vibe-trading options based on hallucinated GPT slop analysis

zuck hired wang for 0.5-1% equity he’s basically a founding engineer

it’s DeepMind > OpenAI > Anthropic > xAI and all of those separations are quite large

the most important equity offer you’ll ever get has an 18 year vest with a 9mo cliff

if you’re building a RAG system in 2025 just build a good search engine backend + let the model query it

oh he’s making a content house. he’s gonna have an army of tiktokers selling the product all day while adding features to it. wow. will certainly be interesting to watch.

11k youtube views on this already is kinda wild to me lol thanks for checkin it out :)

“why aren’t they talking about R1 in the newspaper” the average american is using 4o-mini on chatgpt [dot] com without an account and doesn’t know or care about o1

zuck fumbled hard by not answering the question about the person whose advice he values most. like dude you have a wife lol

is it just me or has o1-preview been degrading lately?

already exists

(yes i am fully aware that these do not work)

he was too early, but he was right. Large Action Model all the way down

been learning a lot about LLMs etc over the past year, organized some of my favorite explainers into a “textbook-shaped” resource guide wish i’d had this at the start, maybe it can useful to others on a similar journey genai-handbook.github.io

RL is hard to explain to people because it doesn't really make any sense without internalizing strong intuitions about everything going on here

we’re hiring for lots of junior roles right now. apprenticeships, basically. can be part-time, remote, international. if you really want to get our attention for these, share a cool env to the hub, <300 LOC. we’ll give full-time offers to top performers after. we can sponsor.

it's simple, really. GPT-4.1 is o3 without reasoning, and GPT-4.1-mini is o4-mini without reasoning. o4-mini-low is GPT-4.1-mini with just a little bit of reasoning. o1 is 4o with reasoning, o1-mini is 4o-mini with a little bit of reasoning, o3-mini is 4o-mini with reasoning that's like better but not necessarily more, and o4 is GPT-4.5 with reasoning.

update: i joined @primeintellect :) cannot describe how excited i am to be joining such an incredible team and mission. there is a dire shortage of labs who are truly embracing open-source research. it’s hard to get the incentives right. you need a business model where open-sourcing your work is positive-sum; being a GPU marketplace is a really good one. prime intellect has been doing incredible work to advance the frontiers of decentralized training and inference, and it is only the beginning. my own goal is to continue along the directions of my recent projects and musings, but bigger, bolder, more real: advancing open research and infrastructure for agentic RL. towards open-source AGI 🚀

Of course that’s your contention. You’re a first-year AI engineer who learned the hard way that RAG is more than vector search. You just finished reading “Building effective agents” and were convinced to migrate from LangChain to Model Complex Protocol, until next month when

the vast majority of american knowledge workers are not very good at working with LLMs and don’t have a great sense of what tasks are within scope of current capabilities. shocking

most things that people call “agents” should really just be called pipelines

none of the big labs have shipped anything resembling lifelong memory. they haven't really even shipped RAG. none of the open-source stuff works well enough to be useful without intense fiddling for very constrained workflows. this is the real bottleneck now, not "intelligence"

being good at ML systems helps you run more experiments. being good at ML theory helps you run less experiments

just a minor version bump. booooring

you can just RL a model to play Wordle

karpathy is basically the michael jordan of ai educators. except for, well, you know

i'm increasingly convinced that "transformative ai" is going to look like an abundance of specialized models for everything from drug design to weather sims to robotics to supply chains, not one agent to rule them all. we're going to need a lot more ai researchers

Roblox CEO David Baszucki has said gaming for a living could replace labor in a post-AI world

today was my last day at @morganstanley. the ML research team there has been a wonderful home for the past 2 years. i’ve learned more than i ever could have imagined about LLMs, markets, responsibility, and how things work in the real world. i’ve made many great friends, and have nothing but positive things to say about the research culture there. it’s a one-of-a-kind lab unlike anything else in finance. but it’s time for a new adventure. will say more very soon about what’s next. incredibly excited to get to work. the arena awaits.

was just in an all-waymo traffic jam

i feel like you shouldn’t call your paper that

if you’re doing a PhD in a quantitative discipline you should print this out and hang it above your desk lkozma.net/inequalities_chea…

wow. Step Mom is going to become even more beautiful

just now learning that all of the cracked ml twitter anons are like 19 years old

TLDR: codex is so good that people kept trying to use it for harder tasks and it didn’t do those as well on those and then people just assumed the model got worse

lmao

what if they added a file explorer and branch navigation and a merge resolver and debuggers and linters and build systems and markdown rendering and multiplexing and a plugin store

once you understand HTTP, you never see the internet the same way

AI isn't actually intelligent. LLMs aren't reasoning. neural nets aren't thinking. computers don't actually compute. algorithms are fake. circuits are made up. logic is a social construct. math is just an illusion

it would take all of the H100s Nvidia made last year to run 498k copies of R1

the year is 2030. python 5.0 now compiles directly to Rython, borrow-checked, zero-cost, no more gc. pip defaults to uv add, and tokio ships with stdlib. the GIL has been replaced with lifetimes, and del wraps std::mem::forget. we still argue about tabs vs spaces.

this is like if everyone’s LK-99 replication attempts actually worked

benchmark idea take a repo with >10k lines of code + high test coverage delete a random file goal: rewrite the file + pass all tests

sometimes i get linkedin DMs from quant recruiters mentioning that a fund is starting a machine learning team and i’m like what do you mean “starting”

it's kind of a "King of England" situation

i first got into agentic programming about 18 years ago

thoughts on karpathy interview: - agree with him on pretty much everything - literal transformative AGI is not imminent - that's fine + doesn't mean it's a bubble - "normal technology" that improves productivity + growth is very valuable - the market already reflects this

i miss gemini-2.5-pro-exp-03-25 so bad :(

guy who quits xAI because of “Ani” and joins MSL to work on “Step Mom”

people stopped working on ARC-AGI because they realized it was too hard

coding on your iphone: - acquire mac mini - set up dotfiles, nvim etc - install ssh client on phone - ssh to mini - npm install -g @anthropic-ai/claude-code - scripts/tools for web search, e2b/docker, cli chat - use git branches + CLAUDE.md aggressively - profit?

$2K per month is almost as expensive as an actual PhD student

gen z boss and o3-mini gen z boss and o3-mini

if you're a later-year PhD student thinking about internships for next summer, my team at Morgan Stanley (ML Research) is hiring! we're a tight-knit team of ~20 researchers working on everything from time-series to RL to diffusion to LLMs, both applications and fundamentals. we have compute, publishing freedom, and an endless supply of hard and interesting problems. feel free to reach out to me with any questions!

i’m sure he knows plenty *about* computers and ai and everything but like he dropped out at 19 to be a business guy and has never publicly been an engineer or researcher at any point in his adult life

it’s called vibe investing

very confused why you’d pick Anthropic of all labs as the one to protest

"why are you so good at posting" i spent the better part of a decade studying algorithmic game theory, recommendation system feedback loops, and information spread in social networks

spending everything on Alexandr Wang

WOW! 🤯 this groundbreaking dataset from Meta’s Chief AI Scientist has revolutionized the way that we understand vision 👀 🚀 is this one of the highest-impact releases of all time?? ⏳🔥 10 crazy examples below: 🧵

one of the following must be true: - anthropic is very efficient at training but terrible at serving inference - sonnet is way larger than V3 - anthropic allocates very few GPUs to inference and serves at 95%+ margins - deepseek out-engineered anthropic for training efficiency

2+ years of MCP experience required

Apple is SO far behind in AI, they’ve barely shipped ANYTHING unrelated, the new iOS and macOS are an absolute pleasure to use, there are so many clever features which smartly integrate info across apps and workflows, things feel automatic and intuitive, can’t imagine switching

sorry to the haters but this is a very compelling sequence of product demos. the big question is they can get a cursor-like market share lead before openai clones it