François Chollet

The 3rd edition of my book Deep Learning with Python is being printed right now, and will be in bookstores within 2 weeks. You can order it now from Amazon or from Manning. This time, we're also releasing the whole thing as a 100% free website. I don't care if it reduces book sales, I think it's the best deep learning intro around, and more people should be able to read it.

Thinking of AI as a productivity booster for prior workflows is the wrong framing. Like all of the previous waves of computerization/softwarization, AI is a tool that lets you do new things in new ways.

If you can learn one thing that's genuinely novel to you, you can learn anything.

We saw our first meaningful jump in the ARC-AGI-3 competition today @tufalabs went from 0.68% > 1.17% My notes: - .68% is the score of the best template (which is why so many people have this score) - I'm guessing 1.17% is a novel approach. This also gives them first signal as to what is working. I expect frequent score increases due to this - Tufa Labs has been a serious contestant with ARC Prize for multiple years now. Very cool to see them continuing with V3 - We have a $25K milestone prize at the end of June for the best oss solution. I wonder if they'll open it up

I'm very wrong most of the time

Whenever an AI tells me I'm absolutely right, my trust in it drops by a bit

When an agent is allowed to decompose a goal into smaller sub-tasks, it frequently suffers from goal drift. Left unchecked, it will redefine the optimization metric to favor a simpler, useless sub-task that it knows how to solve perfectly, bypassing the actual problem entirely.

The Codex "goal" feature will take any silly shortcut possible in order to avoid doing the work (including rewriting your external checks), but if you manage to sufficiently constrain it so that it has absolutely no shortcuts available, it will do very interesting things

Gemini

Gemini

Most human tasks are not Markovian, the optimal next action cannot be determined solely by looking at the current state. It depends heavily on the past trajectory, the original intent, and context constraints. An agent that cannot compress and track its past trajectory with absolute fidelity is maybe 20% as useful as one that can.

This is true of all agents, not just coding agents. Probably the biggest challenge that most companies run into in their agent strategy is getting agents the right constrained context to work with for a task. Too much information or conflicting sources, and the agent can easily draw from the data and produce the wrong result. Conflicting sources of truth for documents, data sources that haven’t been kept up to date, knowledge management systems that rely on tribal knowledge to navigate, and so on. On the other end, of course, too little information and the upside is highly limited of agents in the first place. Thus, a lot of challenges with AI strategies are actually data strategy challenges in disguise. This is why there’s such a significant premium on getting structured and unstructured data environments setup properly so agents can work with information effectively. Critical for any large enterprise adopting agents, and also a clear benefit in some cases to startups that can be designed this way from scratch.

A mental model for working with coding agents is that they're blind squirrels running into a maze and bumping into walls. You must place the walls (verifiable constraints) strategically so that they end up in the general region you want them in.

If you're a R user: the R edition of the book is out now!

Decision making was the bottleneck all along. Productivity is the rate at which you make open-ended decisions, the rate at which you reduce future paths.