Daily Brief

The cheapest configuration they tested was the one sending the most tokens. Across 11 presets run against their open source AI tutor, doing nothing at all to the context beat every compaction technique on recall, cost, and latency at once, and their own production defaults scored worse than leaving the history alone. Prompt caching is why. With 97% of tokens served from cache, and cached tokens up to 50 times cheaper on some APIs, compaction has to shrink a context by more than 50 times before it pays for itself, because rewriting the context invalidates the cache. Louis-François Bouchard's fr

Deno gives its incident response agents read and write access to production Postgres, Kubernetes, ClickHouse, AWS, GitHub, and Slack, and it works. Agents now close incidents that used to wake a human up. Ryan Dahl's problem is what happens when one of those agents gets prompt injected through the support system it is wired into. He grants that Opus refuses to drop the users table no matter how hard you push it, then says the part that matters out loud: security cannot be wishful thinking that a model stays obedient. The agent is untrusted software, so the guard cannot live inside it. Claw Pa

Over 30% of changes now merge with no review at all, and the wait on the ones that do get reviewed is four times what it used to be. Ankit Jain's read is that the debate about when we stop reading code line by line is already over, because we stopped. His sharper point is what replaced it: an AI writes the code, an AI reviews the code, the two go back and forth in a web UI, and a human skims the thread and merges. When AI reviews and nobody reads, he says, we have configured the wrong thing. He is also here to correct his own five layer trust model from a few months earlier, which missed that

Full episode: https://www.youtube.com/watch?v=-RXD4bTuFTo Me on twitter: https://x.com/dwarkesh_sp

Rich Sutton, who helped pioneer reinforcement learning and wrote the seminal AI essay The Bitter Lesson, has now cofounded Oak Lab with his former student Khurram Javed. Their goal: to build agents that continuously learn from their own experience rather than from us. Rich doesn't think he holds a radical view: "I'm not weird. The field is weird." He says all learning is continual, and the field is the one that needed a new name for it. Rich and Khurram argue synthetic data is "a big mistake." Their "big world hypothesis" is that the world is massively more complex than any agent or simulator,

a16z's Joel De La Garza is joined by Nick Warner of Neo and Max Pollard of Cotool to discuss what happens when cybersecurity tools built to defend against humans and malware suddenly have to contend with AI agents. As frontier models become more capable of finding and exploiting vulnerabilities, many of the assumptions underlying traditional security are beginning to break. They explore why guardrails designed to stop AI-powered attackers can also prevent security teams from doing their jobs, why defenders increasingly need access to multiple models, and how agentic software creates an entire

Part of a series of monthly puzzles with MoMath.

An avatar of Teddy Roosevelt holds court in a replica Oval Office, generating video continuously for eight hours with no reset, and a second deployment is being built to run for sixteen. That duration is the hard part. Sidney Primas explains that a real time avatar can only look backward, because the future frames do not exist yet, so every block it generates inherits the errors of the blocks before it and compounds them. LemonSlice trains with an attention mask that enforces this during training rather than discovering it at inference, and collapses roughly 30 denoising steps down to a single

Ten dollars now buys roughly three hours of continuously generated video, and fifty buys fifteen. Keegan McCallum sets that against the room's own habits, since plenty of hands went up for burning that much on coding tokens inside a single hour. His argument is that the interesting axis in generative video stopped being quality a while ago. Put a real time generation next to one that took minutes and the slower clip still has better motion, but it cost on the order of a hundred times more to produce. Helios, the model he serves, is a distillation of a 14 billion parameter open model, and it i

GPU utilization is a lie. It read 100% straight through pretraining while the cluster was nowhere near well used, so Gabriel Jorge Menezes tracks tensor core utilization instead, and watched it climb as training resolution stepped from 128 pixels up to 1024. That is one of several numbers he argues you cannot train at this scale without. InfiniBand counters are exported by nothing off the shelf, and most of their failures turned out to be cross node communication, so they built that collection themselves. Any GPU running hotter than 78 degrees gets pulled rather than debugged, because one warm

Uber could not exist without GPS. Ahmed Ahres uses that to argue real time is a change of medium rather than a speedup: before GPS you consulted a map somebody else had already made, and afterwards your own position became something you could act on continuously. He runs the same argument through film. Once a viewfinder showed you what you were shooting, you could adjust while shooting, and that is the reason Instagram and TikTok were possible at all. Generated video today sits on the wrong side of that line. You prompt, you wait, you get a file back, and there is nothing further to do with i