Browse entries 1417–1440 of 1531. Return to the first page to search and filter the complete collection.
Play video
On a literature review task where every paper already fit inside the context window, adding a memory harness changed nothing: the same accuracy, at higher cost. That negative result is the most useful thing in Stefania Druga's experiment, because it marks the boundary.
Play video
Sonnet 4.5 developed what Anthropic's Applied AI team came to call context anxiety: approaching its context window limit, it would wrap work up early and stop with room to spare. They built context resets into the harness to compensate.
Play video
A newsletter writer walked Matt Dailey through an agentic pipeline good enough to amplify their own voice instead of flattening it, then mentioned they were now effectively writing a book every week. Dailey asked whether the audience was reading a book every week. They were not.
Play video
Superconductor left a meeting bot sitting in a Google Meet at their expo booth for four hours, just listening. Someone passing through said they wanted coding agents to have clear acceptance criteria before declaring work finished. Nobody filed a ticket.
Play video
When Fable was pulled back and access to frontier systems stopped looking guaranteed, Lucas Atkins watched enterprises move to Chinese open models, not because they scored better but because availability could be counted on.
Play video
Once GPT 5.3 Codex Spark started serving a thousand tokens per second on Cerebras, inference stopped being the bottleneck and the network became it.
Play video
Chat and coding assistants still hand you walls of text when a button, a chart, or a small interactive view would say it faster.
Play video
You invoke a tool and expect an answer, but real work takes time, and over that time connections drop, networks blip, and processes crash.
Play video
You cannot tell great engineers what to do, and you increasingly cannot tell what an agent did either, so Vaibhav Gupta's answer is to fight slop with slop.
Play video
Reinforcement learning has been easy to sell where the answer is checkable, like math or code, and Will Brown's talk is about everything else. Most valuable tasks have no clean verifier, so Prime Intellect's work is on how you build reward signal when there is no ground truth waiting.
Play video
In this conversation, Olive Song, who leads reinforcement learning at MiniMax, opens up the stack behind the company's open weight models and the infrastructure that serves them. Her starting point is a belief in open source: put the weights out, let builders optimize on them, and share the capability widely.
Play video
Mahesh Sathiamoorthy's pitch is to stand in the researcher's shoes: the hard part of post-training is not the algorithm but the data and the environments that feed it.
Play video
Thais Castello Branco's starting point is that AI is still badly behind on the subjective work, the writing and design where quality is real but hard to pin down, and that ending the slop means building data and reinforcement environments for taste.
Play video
To train an agent that can run production software, you need training data that looks like production, and that is what Joseph Wang's team at Emulated builds.
Play video
A card gets declined and no one, including the customer, can say exactly why. That gray zone is where Divakar Kumar points his agents. In a payments and fraud system, a rule based engine and an ML model already score most transactions cleanly; the hard cases are the ambiguous ones that neither can resolve.
Play video
Humanity compressed the road from the enlightenment to the moon landing into a few hundred years, and Richard Socher's wager is that automating research compresses it again.
Play video
The slowest part of shipping a production finance agent is not the model or the GPUs, it is you, the developer in the loop. Ramana Siddanth Emani's point is that the same agent harnesses you use to build products can automate your own developer loop.
Play video
Since skills were open sourced, Yogendra Miraje's team at FactSet stopped thinking about shipping features and started thinking about shipping skills.
Play video
Nubank serves 135 million customers, so an AI agent that mishandles a support conversation fails at scale. The talk opens with the result: five agents in production, higher customer satisfaction, and roughly 20 times faster shipping.
Play video
When coding agents got good enough at the end of 2025, Morgan Stanley's roughly thirty person research group asked what would happen if agents ran the research, not just wrote the code. The result is AlphaLab, a multi agent system they built and open sourced.
Play video
Almost every agent today is built for one user, and Sai Krishna Rallabandi has spent about eight months on what breaks when you drop one into a group chat instead. Running a personal agent across a real group, and eventually onto glasses, forces two hard problems.
Play video
Two runs touch the same session, the second write silently erases the first, and the agent keeps answering with total confidence from stale state. Nothing crashed and the model did not hallucinate, so this is a harness failure, the kind that lives in the system around the model rather than in the weights.
Play video
DeepSWE is 113 software engineering tasks written from scratch, not scraped from pull requests, so a model cannot have seen them in training.
Play video
In the world of AI mental health, vibes can be dangerous, with real consequences. Building SonderMind's Mental Health AI Coach required a new playbook for eval-driven development that balances effectiveness and safety.