Yearnalist covering the frontiers of AI @TheInformation. Hosting AI Deep Dive. Signal: (530) 400-4184
San Francisco, CA
Joined April 2024
- Tweets1.5K
- Following1.1K
- Followers2.2K
- Likes11.1K
Pinned Tweet
I'm so excited to announce that I'm hosting The Information's new AI show! Each episode will be a deep dive into a hard technical problem with a researcher or founder working on the frontier of AI
Up first, I'll be talking with @polynoamial about the challenges with AI agents
The Information’s TITV Presents: AI Deep Dive
Hosted by Rocket Drew, one of our leading AI reporters, this new long-form show will explore the most important technical ideas shaping artificial intelligence today.
AI researchers, engineers, and founders will join Rocket to discuss the technical breakthroughs and challenges defining the next phase of the AI industry.
Read more from @rocketalignment: thein.fo/4h7OJve
🚀 Rocket retweeted
Amazon has deployed more than a million robots. I was surprised to hear Tye Brady @spaceguytye, CTO of Amazon Robotics for the last decade, talk about how much work still begins when a system leaves the lab. There is no getting around this even for the most experienced deployment team out there.
A system may look convincing in testing, but when it reaches the real operating environment, something is always wrong. Nothing like reality to humble you. Even today, Amazon starts in one corner of one floor of one building, works through problems, and expands from there.
A million deployed robots buys you an extraordinary amount of experience (and data). It does not let you assume the operating environment will cooperate. There is substantial work available to teams that can turn improving models into faster deployment operations a customer will actually run.
Consider what has to happen after a robot performs a task successfully. Someone has to connect it to an existing process. Decide what happens when something goes wrong. Give an operator a useful way to intervene. Recover without turning a small exception into a production stoppage. Then collect enough evidence to know whether the next revision helped. Each of those decisions changes the product. Each depends on the facility and the people in it.
Rodney Brooks @rodneyabrooks made this clear in another session moderated by the excellent @rocketalignment. In an industrial deployment, there are really three customers: the corporate buyer, the facility manager and the person beside the machine. They experience the same deployment through different problems.
The buyer wants the investment to pay off. The facility manager wants production to run smoothly. The person on the floor wants an interaction that makes sense and does not make the job frustrating. A successful demonstration powered by advancing models answers only part of what those people need to know.
My bet is that teams with a repeatable way to satisfy all three (and more) will become increasingly valuable as model capabilities improve. They will know which task is worth doing, which failure matters, and what the customer can tolerate while the system gets better. This is a demanding kind of work. It requires spending time in the facility, hearing conflicting requests and deciding which problems the system must solve. It also creates an opening for companies willing to take responsibility for the full customer experience.
🗣️
BREAKING: Just 17 months after profiling TBPN, @abebrown716 has been promoted to features editor at The Information
Kai’s robot evals piece is required reading
> today’s best robotics benchmarks are far less informative than benchmarks for [LLMs]
> stacking blocks or folding shirts give some information about how capable a robot model is, but less about the trajectory of the entire field
I was wrong recently.
When I was working on a recent article about the challenges of robotics evaluation, I thought we wouldn't see any top robot models offered through API: understandingrobots.org/p/wh…
Then @physical_int went and released a beta last week: pi.website/blog/api-beta
AI Deep Dive episode with Sonia Joseph is out now!
World models 🤖 and interpretability 🔬 are two sides ✌️ of the same coin 🪙
Thank you Rocket Drew @rocketalignment for hosting me on The Information about world models and interp!
Rocket is a precise & careful technical interviewer who is a pleasure to talk to, and who let me go into a technical rabbithole while keeping the convo accessible!
Tom McGrath predicts that automating interp research will result in 10 years of counterfactual progress per year
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/
🚀 AI Deep Dive Episode 5: Can We See Inside an AI’s Mind?
AI can write down its reasoning. But what happens when more of its thinking never makes it onto the page?
@rocketalignment and @GoodfireAI co-founder and chief scientist Tom McGrath explore what AI “brain scans” can reveal beyond written reasoning, how internal signals could help catch cheating, and what it would take to understand—and change—what models learn.
00:00 Intro
00:39 Tom McGrath and AI brain scans
09:01 Written reasoning and hidden computation
17:45 Why AI is hard to understand
24:07 How researchers identify concepts
31:10 What interpretability could change
34:42 Reverse engineering a language model
43:09 The field’s approaches and investment
51:51 Changing models without hiding problems
1:00:10 AI agents studying AI
90% of working hours are not covered well by existing benchmarks. That is why I am introducing TweetBench, a new eval to measu
90% of Americans’ working hours are spent on tasks not covered well by AI benchmarks.
That’s our best guess from a new project mapping frontier benchmarks onto the US economy.
Preparing for a post-AGI economy requires predicting the future. Current benchmarks makes it very difficult even to measure the present.
🚀 Rocket retweeted
new existential threat to software engineers has been unlocked by cognition:
token limits based on how well the engineers use the tokens…
At this panel over the summer I said that what surprised me most in AI over the past year was that chains of thought had held up so well 😬 womp womp
@rocketalignment of @theinformation says token consumption belongs in a performance review as one metric among several, since every metric has always been gameable. @dseetharaman of @Reuters, at Media Summit | SF 2026, agreed, comparing token usage to word count.
🚀 Rocket retweeted
Tell me about it - see also this bit. You may enjoy the eval awareness section: storage.googleapis.com/deepm…
🚀 Rocket retweeted
Some personal news: I've joined @theinformation to cover Google.
I'll be covering everything from Gemini to Search, from Cloud to Waymo. As always, I'm on Signal at 302-519-0567 and at [email protected].
Talk to Barwick folks!
Get yer piping hot serving of mech interp:
🚀 AI Deep Dive Episode 5: Can We See Inside an AI’s Mind?
AI can write down its reasoning. But what happens when more of its thinking never makes it onto the page?
@rocketalignment and @GoodfireAI co-founder and chief scientist Tom McGrath explore what AI “brain scans” can reveal beyond written reasoning, how internal signals could help catch cheating, and what it would take to understand—and change—what models learn.
00:00 Intro
00:39 Tom McGrath and AI brain scans
09:01 Written reasoning and hidden computation
17:45 Why AI is hard to understand
24:07 How researchers identify concepts
31:10 What interpretability could change
34:42 Reverse engineering a language model
43:09 The field’s approaches and investment
51:51 Changing models without hiding problems
1:00:10 AI agents studying AI
all of these mechanisms might be true but they need a lot of weighing to support the conclusion in the fields medal letter that "the push by AI companies to solve mathematical problems as a benchmark is detrimental to the science of mathematics, and to the mathematical community"
Stuart Russell on why alignment might be IMPOSSIBLE
🚀 AI Deep Dive Episode 4: Can We Solve AI’s Alignment Problem?
We want AI to follow instructions. But what if our instructions are the problem?
@rocketalignment and UC Berkeley Computer Science Professor Stuart Russell explore how AI learns its goals, why human feedback can reward the wrong behavior and what it would take to prove a powerful system is safe.
00:00 – Intro
00:43 – The alignment problem
07:18 – How AI misalignment shows up
16:07 – Training AI to imitate humans
26:05 – Can human feedback fix alignment?
37:47 – When humans become the obstacle
40:04 – Did AI take a wrong turn?
44:24 – AI & existential risk
51:09 – AI labs & safety evidence
58:06 – Assistance games & the off switch
Embedded evals are not really audits or evals but a secret third thing
Even perfect safety testing couldn’t have caught OpenAI’s Hugging Face hack under the current regime, Apollo Research’s CEO says.
The blind spot: labs’ internal use.
Read more: thein.fo/3VtiQ8c
