Building AI-First Companies in the post AGI era. Companies: HotMic, TIDY, SignNow
Costa Mesa, CA
Joined August 2010
- Tweets2K
- Following2.2K
- Followers511
- Likes3.2K
Wow this is really good. Cache expiration was a huge isssue (I think cache forever is unnecessary but say 5 week cache is enough to avoid almost any real cache issues.
Chris Hawkins retweeted
I cannot let a steam engine analogy go by without bringing up John Wilkinson, perhaps the most underappreciated engineer of the Industrial Revolution.
Watt had figured out how to dramatically improve the steam engine, but he spent years struggling to actually manufacture one. The problem was embarrassingly simple: nobody could make a sufficiently round cylinder.
The casting and machining tolerances were atrocious. Watt complained that one 18-inch cylinder was out of round by as much as THREE-EIGHTHS OF AN INCH. Steam leaked past the piston. He tried packing the gaps with cork, oiled rags, paper, and apparently even old hats. Nothing worked particularly well.
Enter John Wilkinson, a British ironmaster whose immediate concern was manufacturing better cannons.
At the time, iron cannons were commonly cast around a core, introducing defects and misalignment that could make the barrels inaccurate or dangerously weak. Wilkinson developed a process for boring cannons from solid castings, and in 1774 patented a vastly improved boring machine.
The key innovation was almost comically mechanical: a massive, rigid boring bar, supported at both ends, that could cut a genuinely straight, circular hole instead of following the irregularities of the casting.
He adapted this technology to Watt's cylinders. By 1776, Matthew Boulton was boasting that Wilkinson had bored a FIFTY-INCH cylinder accurate to within the thickness of an old shilling.
Suddenly, the steam engine was a commercially viable machine.
And the story gets better. Wilkinson became one of Watt's first customers, buying an engine to drive the blast machinery at his own ironworks. The steam engine was now helping manufacture the iron that made more steam engines possible.
I find this whole episode absolutely extraordinary. The Industrial Revolution was, for a critical period, bottlenecked on the ability to make a sufficiently round hole.
We remember Watt, and rightly so. We remember the invention of the steam engine and the subsequent development of thermodynamics. But the man who made the machinery that made the whole thing possible is practically a footnote.
I think we consistently underestimate how much of scientific progress is actually downstream of advances in tooling, measurement, and manufacturing precision. It's not just that technology precedes theory. Sometimes the decisive breakthrough is that someone finally figures out how to build the thing the theorists have been talking about for twenty years.
Anyway, I think John Wilkinson deserves considerably more credit for modern civilization than he gets.
a few thoughts:
1. pretraining science is kind of like how people built quite good steam engines before they understood thermodynamics. “Science owes more to the steam engine than the steam engine owes to science.”-(L.J. Henderson). if nick is right, which i am pretty sure he is, progress points in the direction of more interpretability, and pretraining was "slouching towards monosemanticity". my guess would be that major breakthroughs in interp also advance capabilities
2. nick is a good friend of mine, safe to say my origin story at openai wouldn't have happened if i hadn't met him, and it's fascinating watching him operate and seeing the early drafts evolve into this. he's the most profoundly happy guy I know- not driven by guilt, activism, or religious terror, but by fascination. he of course understands the stakes of understanding neural networks, but it doesn't seem like the core reason he does it: he just loves the neural nets. afaict he's posting this from some beautiful hideaway in tokyo. there's a lesson in there for all the +EV maxers. i feel very grateful to have met him
3. i do think it's crazy how few people across the world were staring at neural nets growing: most have moved to a different level of behaviorist analysis, thinking about reward functions and test scores and whatnot. it does not bode that well for the major model companies, oai included, that fundamental breakthroughs like this are coming from outside. we've become too much of a pressure cooker focusing on local incentives, studying narrow tracks under old paradigms, mostly doing engineering, rather than fundamental research, and implicitly hoping the next model will figure it out
Seems like a big deal if there appears to be a clear path to interpretability -> alignment.
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/
Brutal take…but I do think Google models are spikey and different in interesting ways. I’m a huge fan of having different models check and provide feedback to opus, and I had more unique feedback from Argon. Half was bad, but half good. So agree this isn’t a good coordinator model, but it’s a good model to throw ideas at opus.
Opus 5.5 is so good
Hey, AI give me your best "Unless it is from X, its just sparkling Y" style joke":
GPT-6 Pro: "Unless it comes from the Stockholm region of Sweden, it’s just sparkling employee engagement."
Gemini 4 Argon High: "Unless it comes from the 3:00 AM region of staring at the ceiling, it’s not an existential crisis—it’s just sparkling anxiety."
Opus 5.5 Max: "It’s only cognition if it comes from the Cognac region of France. Otherwise it’s just sparkling autocomplete."
Chris Hawkins retweeted
This is an excellent visualization of Columbus’s voyages!
A good reminder he was sailing into the complete unknown.
Chris Hawkins retweeted
Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel.
- Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all.
- Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract.
- Now we get to professional use of frontier-grade LLMs in math&code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries... This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code.
- And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work. Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the "archeology" of the OpenAI-HF incident from many months ago. Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while.
The funnel is driven by a combination of factors:
1. The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about. This aspect drives the consumer / professional dimension of the funnel.
2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel.
3. Access. Free-tier, paid-tier, internal.
So this is the weirdness of the moment. The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time.
Judging by my tl there is a growing gap in understanding of AI capability.
The first issue I think is around recency and tier of use. I think a lot of people tried the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much. This is a group of reactions laughing at various quirks of the models, hallucinations, etc. Yes I also saw the viral videos of OpenAI's Advanced Voice mode fumbling simple queries like "should I drive or walk to the carwash". The thing is that these free and old/deprecated models don't reflect the capability in the latest round of state of the art agentic models of this year, especially OpenAI Codex and Claude Code.
But that brings me to the second issue. Even if people paid $200/month to use the state of the art models, a lot of the capabilities are relatively "peaky" in highly technical areas. Typical queries around search, writing, advice, etc. are *not* the domain that has made the most noticeable and dramatic strides in capability. Partly, this is due to the technical details of reinforcement learning and its use of verifiable rewards. But partly, it's also because these use cases are not sufficiently prioritized by the companies in their hillclimbing because they don't lead to as much $$$ value. The goldmines are elsewhere, and the focus comes along.
So that brings me to the second group of people, who *both* 1) pay for and use the state of the art frontier agentic models (OpenAI Codex / Claude Code) and 2) do so professionally in technical domains like programming, math and research. This group of people is subject to the highest amount of "AI Psychosis" because the recent improvements in these domains as of this year have been nothing short of staggering. When you hand a computer terminal to one of these models, you can now watch them melt programming problems that you'd normally expect to take days/weeks of work. It's this second group of people that assigns a much greater gravity to the capabilities, their slope, and various cyber-related repercussions.
TLDR the people in these two groups are speaking past each other. It really is simultaneously the case that OpenAI's free and I think slightly orphaned (?) "Advanced Voice Mode" will fumble the dumbest questions in your Instagram's reels and *at the same time*, OpenAI's highest-tier and paid Codex model will go off for 1 hour to coherently restructure an entire code base, or find and exploit vulnerabilities in computer systems. This part really works and has made dramatic strides because 2 properties: 1) these domains offer explicit reward functions that are verifiable meaning they are easily amenable to reinforcement learning training (e.g. unit tests passed yes or no, in contrast to writing, which is much harder to explicitly judge), but also 2) they are a lot more valuable in b2b settings, meaning that the biggest fraction of the team is focused on improving them. So here we are.
I think it’s funny that somebody is bragging about being the top of the leaderboard for more than two weeks. Sign of the times.
Our latest TTS model has been out for two weeks, and it still leads the @voicearena_ai US English latency board.
Sub-100ms median time to first audio, down from 236ms for our previous model.
Try it: gradium.link/studio
Heterogeneous agent architectures seem obviously the way of many things. I already combine software image models jev and LLMs together for many usecases. Cool to see some interesting packaging here!
Introducing Kardashev-0.7, the world’s first trained swarm made of 32 distinct models
Trained with RL for Population Scaling (RLPS), 32 models organically develop specialization & complementary capabilities, delivering frontier performance at:
- 0.007x ~ 0.02x of the inference cost
- 0.03x of the required memory
Civilization advances through different minds specializing and working together. We’re bringing that principle into AI
@BanburyRoadAI, we’re scaling intelligence by model count, toward civilizations of models that learn to build on one another
Chris Hawkins retweeted
A new piece in the Financial Times created a great visualization of the results from this study.
TL;DR: Rent control helps people whose rents get limited, but it hurts everyone else. It pushes up rents, pushes down new building, and makes cities worse.
It is horrible policy.
Chris Hawkins retweeted
It is infuriating realizing how much these mobsters cost us.
If they died off and we got automated ports, we could be so much wealthier, so much more resilient, and so much less at risk of being held hostage by these evil men again.
"Accounting" feels like the first major profession that might be taken over by AI. I believe accountants will be asked to move up to strategy and high trust things, which may earn them more in the long run. But AI is really good at spreadsheets, math, and reading laws. And unlike software, there is no Jevon's Paradox for accounting. People may want a little more...but not that much!
'More striking is how fast AI took the lead. Just eighteen months ago, the best AI models fell short of the average accountant’s ~37% score. Today, models ace those same tasks.'
'These results are provocative. So much so that we considered not publishing them for fear of misinterpretation. But we think transparency about the findings matters as people and institutions prepare for rapidly advancing AI.'
Chris Hawkins retweeted
"Life should not be a journey to the grave with the intention of arriving safely in a pretty and well preserved body, but rather to skid in broadside in a cloud of smoke, thoroughly used up, totally worn out, and loudly proclaiming 'Wow! What a Ride!'"
– Hunter S. Thompson
We will eradicate human mosquito bites someday.
Only a tiny fraction of the world’s mosquito species prey on humans. All of these are invasive across essentially their entire range, having tagged along with humans as we moved around the world. Multiple careful studies say that eliminating these species would have little or no impact on the environment. However, it would save at least a million human lives a year, and get rid of a constant source of discomfort that prevents people from safely spending as much time outside as they would prefer. Malaria is only one of dozens of diseases, many potentially fatal, spread by mosquitoes. Gene drives can safely and relatively inexpensively eliminate these parasitic insects; we can create male mosquitoes that can only have male offspring and which then crash the population. It’s well past time that we did it.
Smart moves plus digitally simulating humans will accelerate treatments and cures for everything. Not only will treatments get to people faster, but more will get done because risk is lower.
How ARPA-H is planning to cut clinical trial times from 10+ years to <4 years:
1. 'Phaseless' clinical trials - move towards replacing phase 1 → 2 → 3 with adaptive trials that can cut failed drugs (or advance good ones) faster.
2. Simulate trial outcomes before starting, with digital twins - to reduce reliance on large control groups.
3. AI agents for trial operations.
4. Turn ordinary hospitals into reusable clinical trial sites, and gets AIs to flag eligible trials to doctors and patients.
I’m a very big believer in going deep with a bit on a particular thread. My life is very much not single threaded, why would I want my bits to be that way?
ben thompson predicts most people and companies will end up with just one agent.
the reasoning is that an agent gets better the more it knows about you and the more of your accounts it can access, and that makes it sticky.
i use at least 5 agents every day (claude, chat, hermes, muse, and qm), and i don't think that's how this plays out.
first, they're each better at different things, in the same way that no single app is the best at everything. "one agent" assumes one company will be best at everything. which is a pretty strong assumption.
second, the lock-in is much weaker than it sounds. a new agent is missing basically two things the old one has: your logins and your context.
1. logins. giving a new agent access to my accounts takes about 10 minutes. it can pull cookies to reuse the browser sessions i'm already signed into, or if it can connect to 1password, then it can pull any password. and that doesn't take into account the fact that every website and app is pushing hard on making it easier for agents to sign in.
2. context. most of what an agent knows about me lives in files it wrote itself (notes, preferences, past tasks). this is just MD files in a folder. moving my context to a new agent is copying a folder.
so what does the switching cost actually look like?
a lot more like switching between chatgpt and claude than switching from iphone to android. plenty of people already flip between models every few months when a better one comes out, and i'd expect the same with agents.
of course it's a huge prize, but i think it gets split between ~5 winners, each of which is best at something in particular.
we'll see how sticky each one turns out to be!
Initial thoughts on OpenAI Dev Day: Price increases, a minor model bump, and some experiments which COULD get cool with the most interesting being Dots + Spaces combined (not stand alone) if they evolve that well.
Itemized
1. Dots - Single threaded with a beefier model vs Muse/Grok Bot. IMO multiple agents is critical for my usecases, I don't like single threaded. Although Muse has amazing speed and good enough intelligence. I would be curious why I would need mega intelligence in single thread. Interesting but not obviously a slam dunk.
2. GPT 6.1 Sol - Seems great, odd they update after 1 week but Sol is a great daily driver (not as good as Opus 5.5).
3. Subscription Pricing Increases - Basically, charging $500 for what used to cost $200. Sucks, not shocking, but it does mean that my $200 plan which runs out each week won't cut it at my current model usage.
4. Subscription usage in other products - Not useful to me but cool.
5. Cloud stuff - Cloud agents are overhyped IMO. They have been more irritating to work with and more expensive with no gain for me. Cheaper to buy a cheap server and run all your stuff there.
6. Plugins - These may really make agents more useful. I don't need them because I build agents crafted as standalone repos with tons of software to support them. But for less intense work easier deployment, its nice.
7. ChatGPT Space - Should have been a slack competitor first and foremost with these doc collab second. If they iterate here, this plus dots could get really interesting.
I like the “decisions model” term more than “system one” label jev used.
These really help with large scale bot orchestration (which can be very token burning).
