Jacob Effron
Related people
Podcasts
Episodes
Notes
Unsupervised Learning with Jacob Effron · Ep 97: Jev Creator on Decision Models, Benchmark Fatigue & Why the…
-
Why Is True AI Agency Two Breakthroughs Away?
Diogo Almeida guesses that true AI agency is something like two gigantic breakthroughs away. Reading a chain of thought and calling it agency does not close the gap. He guesses that even one breakthrough would not be enough, though one gigantic breakthrough might. A human who has super tools and wants to do damage is a different threat from ASI or from RSI.
A lot of RSI talk, in his view, skips the history of AI and Amdahl's law. Putting every kind of intelligence on one line, then placing humans and chimpanzees on it, looks dire and does not match the systems. He would not count a 10% speedup on AI research as RSI unless that speedup itself kept rising. He is open to evidence and treats th
-
Why Does Diogo Almeida Call Financial Engineering the Labs' Strength?
Diogo Almeida's sense is that the labs' greatest strength is financial engineering, not models, product, or distribution. As far as he has heard, veterans talk about structuring deals so the lab is protected while other people take the risk of latching onto the rocket ship and accepting the terms. Asked whether that is about compute, he says it shows up on all sorts of sides, and especially on compute.
He does not know how tangled the deals are, or how exposed a lab would be if the market turned. A bubble can pop because expectations are missed, even when the technology is transformative. He has no private information, tries to stay calibrated, and calls himself the more conservative person
-
Why Isn't AI Automating the Easy Work Yet?
Diogo Almeida says models have been smart enough for easy automation for years, and that intelligence still is not being put to work. A document he wrote at OpenAI was titled "we could have had AGI a year ago," after OpenAI's definition of doing most of the world's economically valuable work. He reads the spirit of that line as work by volume. Models from the GPT-3.5 era, and possibly the GPT-3 era, could have met it. They were already past a median level on all sorts of tasks.
That should have meant tons of automation, at least on the easy work. GPQA, the hard multiple-choice test he describes as PhD-level, was passing that bar in 2023. He calls it tragic that intelligence at that level wa
-
Why Is a New AI Model Shape Cheaper Than Reliability?
Diogo Almeida says a new model shape is cheap and reliability is the part that matters. He still wants novel shapes and plans to keep shipping them, but people in machine learning are over-focused on the shape of the task. When he thought the work would take a week, he meant a week to make the shape, and that part did take a week. Other teams are now making similar shapes. The shape does not matter next to the model being useful.
He expects many more shapes than the two already out, even if Jev ships them alone. Too many shapes without utility burn goodwill, because the releases over-promise and under-deliver. He does not want creative energy spent on things that clearly do not work. He is
-
Why Can AI Solve Hard Math but Not Accounting?
Diogo Almeida says the models people need are System 1 models, and the honest move is to say where they fail. System 1 was an unpopular name, because that kind of thinking is often used as an insult for the instinctive, mistake-making side. The shape AI actually gets compressed into is still System 1, and naming the weaknesses is useful for software by itself. Most of what people need is that kind of work.
It should be strange that models can take on hard problems in math and still not automate anything that touches accounting. Cost, speed, and the pitch that a model is good for really fast decisions are distractions beside that north star.
-
How Much Do Coding Agents Speed Up Engineers?
Diogo Almeida estimates coding agents speed up engineering by 20% to 40% on average, with error bars for long-term effects like skill degradation and tech debt. He treats the average gain as a decent but small boost.
The number depends on the use case and on the skill of the user. Side effects are the bad part, so the gain also depends on whether side effects can be isolated in a way that stays legible. In some use cases the speed-up is kind of like an infinite percent, because the work would never get done otherwise. In other use cases the agents are as bad as coding the thing by hand.
A decent but small average boost is the consensus he has seen in the field right now.
-
Does Jevons Paradox Already Hold for AI?
Diogo Almeida says people at Jev truly believe in Jevons Paradox, which is why the company carries the name, and the paradox does not hold yet. Rebound is already visible. Cheap goods create more demand, and every good sees rebound. Backfire is different: the total market size increases. Jevons Paradox holds in the backfire case, and that case is absolutely not here.
For the paradox to arrive, a key part has to be in place. The product has to be so good and so cheap that a user throws it all over the place because it makes value across place after place. Then the budget for AI actually increases. Rebound counts as settled, and backfire is still absent.
-
How Far Along Is Jev's Image Model?
Diogo Almeida says Jev's image model is maybe 10 percent of the way the text model is. He would be truly embarrassed to call it production ready, because so much jaggedness is left, and it still lacks his seal of approval.
Text tasks will be automated before image tasks. Investment went really heavily into text tasks, with a tiny amount of work on image tasks by comparison. He is jazzed the image model does as well as it does. The lead in images is not as large as the lead in text, but he senses an overall lead in images and could be totally wrong. He calls the image model completely insufficient. It is not quite as bad as other people just shipping a shape of something. In the grand scheme
-
Why Is Key-value Cache Compaction Hard to Solve?
Diogo Almeida says compaction fits a write-once, append-only key-value cache, because a full cache gets restarted, but compaction is compression and compression is not fully solvable. Compressing without priors is very hard. Compress for the problem just solved, then ask a different problem that needed information already thrown away, and the agent is dealt a losing hand. He does not know the alternative. He calls it science fiction: notice a new task, let code compute the expected value of cache reads and cache writes, rebuild a cache for that task, and keep the old cache hot.
Working the same code base means rereading the same files and relearning the same abstractions. Markdown files wri
-
Was Diogo Almeida Mission Driven at OpenAI?
Diogo Almeida would have called himself one of the lowest percentile mission-driven people at OpenAI. The OpenAI mission was left vague on purpose so many people's missions could fit, and he never really believed the grand mission, even while wanting AI to be extremely transformative. He is still unsure whether he was mission driven then. He is mission driven now.
Early confidence came from a lot of work spent squaring two facts at once: AI is super incredibly awesome, and AI is really not useful in the productivity stats. Most people focus on one side or the other and never land a sensible bridge between them. Distribution is a nonsensical answer. So many companies have gone all in on dist