AI Consultant 789Labs | Ex-Microsoft AI 🧑‍💻 🌉 👾
Joined November 2014
- Tweets1.8K
- Following381
- Followers152
- Likes3.6K
Julien retweeted
Be extremely intentional with your friendships — not acquaintances, not business partners, actual friends
Try to pick high-integrity people — the right people
Invest time in those relationships
Cut or downgrade folks who are transactional or who don’t show up for you
Reallocate that time and energy to the right people who do
It's ok to be cutthroat about this, as life is the memories you create with your friends and family.
Invest in making moments.
Julien retweeted
Why are we calling every new kernel release a Megakernel. This is a well written CUDA graph engine with persistent kernels and PDL.
The author didn’t even call this a megakernel
github.com/vllm-project/vllm…
BREAKING NEWS🚨: NVIDIA has just achieved 423 tok/s/user 🚀 on the 2.8T Parameter Kimi K3 model via a new upstream vLLM optimization. 1.68x Faster on the Same Hardware!
This was done by NVIDIA CentML's Benjamin, who wrote NVIDIA's new decode-specialized kernels in Cute DSL via aggressive Megakernel fusion. This is a 12,000 LoC Megakernel and is the future of advanced AI-assisted kernel generation. NVIDIA is already working on DeepSeekv4.1 and other models for vLLM too.
Someone should give this NVIDIA 10x engineer a raise for this great new major feature!
Julien retweeted
When making money is no longer your goal, that is when you can start being truly effective.
Julien retweeted
Every person that moves to SF goes to a bunch of tech events in the first two weeks and realizes it's a complete waste of time.
Julien retweeted
Some observations about the "total" automation of software engineering:
1. There are more software engineers now than in 2025, and demand is growing.
2. Software engineers have never worked harder than they do now.
3. They largely don't write code anymore, agents write ~100% of the code.
4. The bottom line of the companies that employs these engineers and agents (i.e. the terminal purpose of all this code) is typically not budging as a result of AI adoption.
This is why you have to think in terms of systems. Automating one aspect of a system creates its own friction and moves the bottlenecks elsewhere.
Julien retweeted
I’m literally that weirdo who wrote (all by hand) a 15K word blog post on:
- why one cant treat models as oracles
- how models gaslight you all the time
- what it takes to build confidence in the model’s output
copyconstruct.medium.com/how…
So yeah sure, I believe in magical woo-woo
Julien retweeted
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
Julien retweeted
I'm simply not going to have it output 100k loc
Code is literally useless. Code in fact is a liability. Having more code is actually worse overall. I am trying to solve problems with the least amount of code possible. This is very, very important
Julien retweeted
Character AI has gone bankrupt.
The AI startup announced its Chapter 11 filing through a late-night newswire press release
Julien retweeted
Introducing Microsoft-Decision-1, our new model for fast decision-making.
It delivers top performance on structured decision tasks, outperforming both LLMs and other decision models in latency and quality.
We’re already testing it across Microsoft for everything from incident response and quality control to scientific discovery.
Julien retweeted
These are some very important observations from Satya.
As models and harnesses commoditize, intent, verification and evidence will gate how quickly SI permeates enterprises and businesses.
Importantly, you need to implement systems that separate who does the work from those who govern, measure and sign off on it.
This is a new kind of “harness” that will emerge over the next few years. It may become a new system of record over time. One whose entire role is to track “change” and map it back to SOPs, policies, laws and the like.
Julien retweeted
Thank you, @POTUS, for this incredible honor today. I'm humbled to stand alongside so many giants of American innovation and looking forward for what we can continue to do together to advance technology and drive American prosperity.
.@POTUS: "For more than three decades, @satyanadella has been a key figure in the transformation of Microsoft — revolutionizing the company and empowering businesses, students, and consumers from all over the world with some of the most capable software ever built."
Julien retweeted
ok it's official... AI-written unit tests and integration tests are empirically proven to be unhelpful. you should tell your agents to stop writing tests by themselves
on the deepswe eval set, banning sonnet 5.5 high from writing any tests actually resulted in slightly higher success rate (non stat-sig), with less time and token spent (stat-sig). this is pretty hard empirical evidence
across the tests that were written by the "tests allowed" baseline arm, 65% of them were unit tests, 35% were integration tests, neither bucket resulted in any improvement compared to not writing any tests at all
i also picked a random sample of 44 tasks subset where i completely disabled executing even existing tests - it also did not affect success rate at all
i spot checked many tests written in the baseline arm, and my intuition is that most tests are simply a repetition of the implementation
agent-written tests do not add any value because both the implementation and the tests were simply the agent's interpretation of our intent. the tests aren't any more accurate than the implementation itself
i suspect we can still extract some value from unit tests and integration tests if we describe them ourselves when we believe we can articulate our intent better through test cases than through requirements. but i have not proven this yet
also worth noting, during deepswe eval the agent wrote almost no e2e tests (only 17, compared to 3000+ unit/integration tests written). so this analysis does not prove nor disprove the value of e2e tests - i will do another eval specifically for that
Julien retweeted
What if the jagged frontier is mainly math + code (which you can push arbitrarily far with RLVR), and everything else starts to plateau because it is still bottlenecked by human generated data?
Model performance in non-verifiable areas has kept improving steadily, albeit much slower than for math and code. But is that steady improvement a side effect of a higher G (itself driven by RLVR), or only a function of the amount of new human data getting injected into training (which is still continually happening on a massive scale)?
A lot of things depend on the answer to this question
Julien retweeted
Hey @ConwayResearch, the fair thing would have been: add @PrismML’s NOTICE file to your releases, tag Bonsai 2 and Qwen3.8 as base models, and credit @inco_ai for Splash, the engine behind @underdogai’s speed numbers.
tldr; give credit where it's due
Underdog’s Ternary is a 27B ternary-weight model from Conway Research (Underdog) for efficient private on-device inference. It derives from PrismML’s open-source Ternary Bonsai 2 (near-lossless ternary quantization of Qwen3.8 27B), with post-training applied. Model card and independent checks confirm the base; sizes and kernels match closely.
Julien retweeted
He’s right.
Microsoft CEO Satya Nadella wrote that a company renting an AI model “essentially pays for intelligence twice”:
Once in money, and again in the proprietary knowledge it has to reveal to make the model useful.
This mechanism is called the learning loop: every prompt and correction captures some of the company's know-how.
That knowledge is distilled and, with a closed model, the provider keeps that information.
If your usage data is worth more than the token cost to serve, providers might end up paying you to use their models. The precedent exists: OpenAI created an opt-in program that gives accounts up to 1M free tokens a day on certain models in exchange for training rights on that traffic.
Chips can be rented and algorithms can be leaked or distilled, but the prompts and corrections flowing through a model compound privately for whoever holds them.
Julien retweeted
Internal AI Usage Cap Memo:
In response to requests for increased AI budgets, they aren’t coming. And here’s why:
We currently offer 16x token spend per employee compared to the average of tech forward companies (Based on Ramp's spend data from 70,000 businesses). I say this to establish - for those of you who request increases, it is unlikely you will get it. Unlimited budgets create unlimited waste - a luxury we do not have. So, budget and use your tokens only on projects that have a clear revenue producing or cost savings. Otherwise, it's likely to be filed under "shit that doesn't matter and not worth spending money on."
To help you, we will be offering training(s) on:
How to figure out what's worth working on (biggest miss #1)
How to use fewer tokens to get same output (think of this like AI hygiene)
How to use the time AI saves you to be more effective
AI is a tool that creates leverage. It helps you get out more from what you put in. That said, company revenue is flat (relative terms) compared to AI usage increases. This suggests that people are (largely) using it on projects that don't matter, or are using it on things that do matter, then spending the saved time working on things that don't matter.
This discrepancy (difference between money we spend and money we make from AI) hints at a skill deficiency in the team. One which I hope we will train up - just as well as we trained up on AI usage.
So now that everyone is using it - it's time to use it well.
Let's crush Q4.
Alex
PS - The more clearly you can articulate how what you do generates a return, the more likely tokens will be in your future. And the way to maximize a return is to show HOW LITTLE you can use and HOW MUCH you can get from it.
