Conversations at the frontier šļøš®, Investing on the side.
San Francisco, CA
- Tweets 653
- Following 1,000
- Followers 15,974
- Likes 4,083
Pinned Tweet
We still have no idea what the human brain is truly capable of.
Reed Jobs shares one of the wildest neuroscience stories Iāve ever heard.
65
181
3,302
1,426,362
Sabrina Halper retweeted
Today weāre launching micro1ās PII transformation model, flow-transform 1.0, delivering frontier-level performance across detection, identity synthesis, and transformation of personally identifiable information.
On PrivacyBench, our model reaches 96.0% F1, outperforming every detection baseline we tested, including Tonic Textual, Claude Opus 4.8, Sonnet 4.6, Microsoft Presidio, Haiku 4.5 and GLiNER2.
Some of the most valuable training data for frontier AI models lives inside fully functioning companies. It captures years of real work across decisions, communications, tools, handoffs, exceptions and the relationships connecting them.
The problem is that this data is also full of PII.
Traditional redaction makes the data safe, but it also destroys the very workflows and relationships frontier models need to learn from.
flow-transform 1.0 solves this by turning enterprise operational data into high-fidelity training data for frontier models by replacing real-world identities without flattening the reality the data captures.
90
111
668
242,017
.@daveholtz is a professor at Columbia, on leave for the year to lead AGI Economics research at OpenAI.
A study he conducted proved that while models are getting better rapidly, but we wonāt automatically capture those gains unless we learn how to adapt to and get the most out of each new iteration. It's an interesting and important problem for the labs to work on.
*For tasks where youāre steering AI toward a specific outcome, 50% of the performance improvement from a better model came from humans changing how they used it, not just the model getting better. "In other words, if we gave people DALL-E 3 instead of DALL-E 2 and said, 'Hey, don't change anything about how you use the model, but hey, it's, it's a better model underneath,' they would've only gotten 50% of those gains."
For open ended tasks, like brainstorming or creating something without a specific target, the gains came mostly from the better model itself.
3
5
67
7,874
Today @FactoryAI announced its latest round - $200M at $5B. A rocket ship.
But I asked @matanSF who was there during the hardest periods of building Factory, and who heās relied on most for advice along the way:
@shaunmmaguire @sequoia @chrisdegnan @krisfredrickson @TheChainsmokers @IvankaTrump @CharltonJBoyd
We have raised $200M at a $5B valuation to scale self-improving software development in the enterprise.
@FactoryAI has grown to serve hundreds of thousands of developers at companies including RBC, Adobe, Nvidia, T-Mobile, and Palo Alto Networks.
We will use this capital to accelerate our investments in research, product, and global go-to-market.
7
8
144
18,439
Sabrina Halper retweeted
I really enjoyed this conversation with @SabrinaHalper a few weeks ago, during which we basically did a lightning tour of all of the AI research Iāve done (along with an all-star roster of coauthors) over the past few years.
More generally, I highly recommend Sabrinaās podcast ā sheās an excellent interviewer, and sheās had some amazing guests recently. Iām still not sure how I managed to sneak my way into the lineup, but it was a real pleasure!
"I think the coolest dataset is that we're actually able to look at how people are using Codex inside of OpenAI. If you look at the cumulative amount of time that people have Codex agents working, the most frontier people inside of OpenAI, their Codex agents are working, I believe, over 70 hours a day. So it really is this force multiplier on what you're able to do. And so if we think about traditional organizational structures where many organizations are sort of pyramid shaped, where you have the CEO at the top and then it gets wider and wider as you go down. Maybe that structure doesn't make as much sense when everybody can delegate a bunch of work to numerous agents that are working below them." - @daveholtz
Most of what you hear about AI's impact is a projection. David Holtz has the data. Heās behind some of the most interesting studies of AIās actual impact on work and business today. And the results are surprising. His work cuts through the noise around AI to a key question: what is AI actually doing to people, companies, and the economy today?
David is a professor at Columbia, currently on leave to lead AGI economics research at OpenAI. We recorded this before he started there.
5
25
3,352
Sabrina Halper retweeted
An update: Iāll be on leave from Columbia for the 2026ā2027 academic year to work full-time with the economic research team at @OpenAI, where Iāll be leading AGI economics research.
Recent events have underscored how quickly the AI capability frontier is expanding. Iām excited to explore how economics can help us understand and shape the development of increasingly capable AI, and how the economy will evolve alongside it.
I canāt think of a better group to work on these questions with than @RonnieChatterji and the rest of the team!
42
25
758
105,805
"I think the coolest dataset is that we're actually able to look at how people are using Codex inside of OpenAI. If you look at the cumulative amount of time that people have Codex agents working, the most frontier people inside of OpenAI, their Codex agents are working, I believe, over 70 hours a day. So it really is this force multiplier on what you're able to do. And so if we think about traditional organizational structures where many organizations are sort of pyramid shaped, where you have the CEO at the top and then it gets wider and wider as you go down. Maybe that structure doesn't make as much sense when everybody can delegate a bunch of work to numerous agents that are working below them." - @daveholtz
Most of what you hear about AI's impact is a projection. David Holtz has the data. Heās behind some of the most interesting studies of AIās actual impact on work and business today. And the results are surprising. His work cuts through the noise around AI to a key question: what is AI actually doing to people, companies, and the economy today?
David is a professor at Columbia, currently on leave to lead AGI economics research at OpenAI. We recorded this before he started there.
3
14
5,032
.@moxie marlinspike: "I think that LLMs are one of the first technologies to actively invite confession. And so the idea behind Confer is to do something similar to Signal of building something that allows people to avail themselves of the utility and awesomeness of LLMs, but to do it in a private way. Itās an end to end encrypted version of AI chat where you can interact with AI chat like you would with chatGPT. But all your conversations and conversation history are private and only, so only you have access to them. Itās using open source models.ā
Moxie was hired by Meta to integrate Conferās privacy technology in their new AI products like Muse, a decade after Moxie worked with Meta to bring the Signal Protocol to WhatsApp.
*Recorded in January, in the middle of the ocean, from a floating sauna Moxie built himself.
9
14
235
22,912
Sabrina Halper retweeted
The internet is the dominant media of our times.
Who dominates the internet, dominates culture, discourse, spending decisions.
The internet is vast, noisy and fast so knowing who dominates it every week gets lost in translation especially as we all have different algorithms.
In the last couple weeks I built an index that works as a truth signal for the internet.
Who won the internet in their category in the past week?
launchindex.app
11
4
28
1,173
"There used to be a moat in software of 'we know how to build this and no one else does, so that's why we're going to win.' That doesn't exist. There is nothing that nobody else can build... But that means the biggest differentiator is being willing to say no, and deciding what are the things that we do build and make incredible, and what are the things that might be passing fads but we say no to." - @FactoryAI CEO Matan Grinberg @matanSF
1
6
57
4,526
Weāre entering a crazy period of cyber risk and Iāve been wanting to do an episode on this for a while.
This week Iām sitting down with the founder of @wiz_io, a cyber company acquired by Google for $32B, to go deep on what happens next. Live from Tel Aviv.
What do you want to know?
Cyber is having a moment
Across 21 major software companies, including Apple, AWS, Microsoft, and Google:
- Reported critical vulnerabilities never cleared 100 per month in four years
- Since spring they've jumped to over 600 per month
Charts of the Week: a16z.news/p/chart-of-the-weeā¦
4
1
59
7,552
"I think half the co-founders of Anthropic were theoretical physicists. Jared Kaplan was exactly in the same area of physics as I was. Dario was in physics back in the day, also at Princeton, but he was more on the bio side. There's been an exodus of physicists to more interesting problems, and I think that's for a good reason"
Why @matanSF dropped out of his physics PhD at Princeton
"The mathematics of string theory is probably one of the most beautiful artifacts humanity has created. In order to even produce something in theoretical physics, there's hundreds of years of literature that you need to catch up on, and also every day there's, 100 new papers that come out. And before you can contribute to theoretical physics, you need to know basically all of math, like algebraic topology, differential equations, algebraic geometry, and these are all things that people will spend their careers going in on, and yet to do the physics, you need to at least have, a pretty deep understanding of these things. And so it really teaches you how to have, good judgment with minimal information."
@matanSF studied physics for a decade, eventually dropping out of his PhD focused on string theory to found @FactoryAI.
14
72
612
287,419
"The mathematics of string theory is probably one of the most beautiful artifacts humanity has created. In order to even produce something in theoretical physics, there's hundreds of years of literature that you need to catch up on, and also every day there's, 100 new papers that come out. And before you can contribute to theoretical physics, you need to know basically all of math, like algebraic topology, differential equations, algebraic geometry, and these are all things that people will spend their careers going in on, and yet to do the physics, you need to at least have, a pretty deep understanding of these things. And so it really teaches you how to have, good judgment with minimal information."
@matanSF studied physics for a decade, eventually dropping out of his PhD focused on string theory to found @FactoryAI.
45
63
512
239,428
Matan Grinberg (@matanSF) of @FactoryAI says recursive self-improvement might never be as good as human-assisted improvement of models -
"People talk a lot about recursive self-improvement. But by definition, recursive self-improvement is just getting to a point where the models can quite literally improve themselves. But improve themselves is not synonymous with hyperbolic growth. The people that are most AGI-pilled, like "AGI 2027, everything's gonna change next year". Sometimes they make jumps to oh, we get RSI, and they assume that that RSI is the crazy takeoff RSI. It It could be the case that we get to RSI and it's just never as good as human-assisted improvement on these things, or it could just be really, really slow to actually get to that point."
2
3
93
9,726
New episode with Matan Grinberg (@matanSF), CEO & co-founder of @FactoryAI !!
We get into so many good topics: Studying physics for a decade, how to measure ROI of AI spend, who to hire in a world where we don't code, policy, the risk of Chinese open models vs. closed-model monopoly, what he thinks about recent security incidents from large labs, and why he believes weāre already living in a post-AGI world.
------
TIMESTAMPS:
(00:00) Intro
(02:00) Young Matan
(5:30) String Theory
(11:10) Leaving academia, becoming cultured, and getting nerd-sniped by code generation
(16:53) Cold-emailing @shaunmmaguire and founding Factory
(18:27) The desert years, and what made autonomous engineering work
(20:55) The new bottleneck is deciding what to do
(23:30) The three phases of enterprise AI adoption, and the end of token maxing
(28:18) Chinese open models, and the monopoly to end all monopolies
(30:30) Sleeper agents, poisoned code, and the labs cybersecurity incidents
(34:30) The White House open-weights policy, @mkratsios47, and the distillation problem
(37:30) The coming billion-dollar agent incident
(42:30) "We're already post-AGI" and AI taking on math
(49:30) Building your own software
(52:20) Reflections & best advisors
9
6
83
6,439