Latest / UpNext AI / Legora’s Legal AI, Weight-Space Programming, and WebHarbor’s Browser-Agent Tests
Transcript
- Welcome to the UpNext AI podcast. It's Tuesday, November 3rd, 2026, and here's what matters in AI today. Legal AI has to work on documents its builders may never be allowed to see. In a Latent Space interview published Monday, Legora chief technology officer Jacob Lauritzen described how that constraint shapes the company’s agents for lawyers. Legora cannot read customers’ confidential uploads or the usual traces of an agent working through them. Lauritzen says engineers and lawyers instead recreate legal tasks for evaluation. One simulated due-diligence exercise gives an agent a data room containing ten thousand files and checks whether its report catches the relevant red flags. The team also groups results by practice area and task, because a new model can improve one kind of legal work while getting worse at another. The product’s Tabular Review turns a batch of documents into rows and questions about those documents into columns. Lauritzen says it supports up to a million cells, but starts with ten rows so a lawyer can refine the questions before running the rest. That small preview is consequential when the full job is a large contract review. There is a business-model question alongside the engineering one. Lauritzen says routine work at some U.S. firms is shifting toward flat project fees, changing the incentive to use a tool that saves time. He says Legora is not seeing the disappearance of junior lawyers; in his account, productive juniors can do more with these tools. That is a practitioner’s view of a changing market, and a concrete example of why legal AI has to fit both confidential workflows and the firms paying for them. From evaluating agents to changing models themselves: a research writeup published Monday asks whether a language model can inspect and edit the numbers inside another neural network. Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab call this weight-space programming. Their approach uses a model’s reasoning to read or change weights directly, without gradient descent or a training loop. In small demonstrations, the authors report identifying what game a network had learned from its weights and writing game-playing networks without environment feedback. Other experiments used tools and forward-pass feedback, so these are not all the same kind of test. Their most striking language-model result concerns Qwen: the authors report that an edit reducing repetitive reasoning loops raised its score on the MATH-500 test from 7.9 percent correct to 48.1 percent correct. That result reflects a repair to looping behavior, rather than proof that the model learned new mathematics. The team did not establish a math improvement on a separate model that rarely looped. These are early, author-reported results, but they suggest a possible way to inspect model behavior and target a specific failure inside its weights. Another boundary for AI builders is what their tools will agree to make. Ars Technica reports that Anthropic added a Claude Code instruction allowing original games and games within a genre while disallowing recreations of specific existing games, even when renamed. XDA verified the rule in a Claude Code release from late October. Ars argues the instruction may help Anthropic manage its own exposure, but a rule in one company’s tool is unlikely to slow the wider wave of AI-assisted game copying. Testing agents raises a related question of control: can you tell whether an agent improved when the website it uses keeps changing? An arXiv preprint posted Monday introduces WebHarbor, a collection of executable website environments for agents that operate through a graphical interface. The researchers built 153 website environments and 3,518 tasks, including multi-step work after sign-in. They describe a process in which coding agents implement task-driven changes and human contributors check visual accuracy, functionality, and task validity. That gives researchers repeatable interactions while retaining more of the complexity of real web services. The paper also reports training a small Qwen-based graphical-interface agent in those environments. On four benchmarks, it generalized to unseen tasks and performed competitively with leading open-source agents. That is the authors’ reported result from a preprint, rather than a guarantee for agents on live websites. For teams comparing browser agents, WebHarbor’s central contribution is a more consistent place to test them. ...Are you building apps with voice? Elevate your app's voice capabilities with ElevenLabs. Their API is a game changer for embedding dynamic, responsive voice interactions in your applications, providing unprecedented realism, flexibility and latency. Visit up next dot fm slash eleven to check out their latest offerings. ... A few shorter developments before we close. Simon Willison has released version zero point three seven of his LLM command-line tool and Python library. It adds an easier way to configure models on OpenAI-compatible endpoints and switches the unset default model to one Willison says costs half as much as the previous default. The Decoder reports that OpenAI has quadrupled bug-bounty payouts, with rewards of up to $172,000 for bug hunters, and launched an internal security sprint. The higher ceiling changes the incentive for outside researchers to report qualifying vulnerabilities. A quick disclosure: we use models from OpenAI to help write the UpNext AI podcast. The Information reports that chip designer Marvell has agreed to acquire laser maker Lumentum in an all-stock deal. The combination brings together two suppliers of technology used around data centers as AI infrastructure spending grows. And Oracle has filed to raise $20 billion through debt and stock offerings, according to The Information. The filing divides that amount between $10 billion in investment-grade bonds and $10 billion in common stock as Oracle spends heavily on AI infrastructure. Before we wrap up, a quick note: this podcast is generated with the assistance of AI and is intended for informational purposes only. All referenced articles, research, and commentary remain the property of their original authors and publishers. And that's your briefing for today. If you enjoyed this episode, share it with a friend or colleague—and follow UpNext AI on Apple Podcasts, Spotify, or wherever you get your podcasts! Full source links are in the episode notes, and we'll be back tomorrow with what's up next!