# Simon Willison

Najpopularniejsze odcinki podcastów
Zaktualizowano:

Demonstrating How Locally-Running Qwen 3.8 27B Identifies Games from Model Weights

Model weights are a massive sequence of floating-point numbers that make up a neural network. To a human, they're meaningless. Just a bunch of numbers. Yet an AI has appeared that can look at those numbers alone and determine this is a model that learned Pac-Man. And it's running on a laptop. Behi…

SUMO Weekly: 31 October–6 November 2026

About this episode An independently researched digest of worthwhile A.I. developments from the week. How this episode was made SUMO is researched and created with the help of ChatGPT. Ben, our real human in the loop, sets the editorial direction. Sources reviewed The complete list shows material SU…

#40: Contain first, then fix

Top Story: GPT-6 Cyber brings stronger findings and candidate fixes — OpenAI introduced GPT-6 Cyber on November 5 for selected approved Daybreak Red users. WorkOS's incident now has a more specific explanation. — WorkOS says independent forensic analysis validated a researcher's account linking the…

2026-10-29 | OPC Brief (15 highlights)

【KrsNews】2026-10-29 OPC Brief 15 hand-picked stories—OPC Tools 8 · Industry Dynamics 3 · In-Depth Report 3 · KOL Perspectives 1. Episode Story Links: [OPC Tools] NVIDIA Open-Sources Nemotron 3.5 Speech Model with Granary Adapters for Complex Audio TasksOriginal Link: https://huggingface.co/blog/n…

NVIDIA and Korea’s Sovereign AI Push, Anthropic’s Power Deal, and OpenAI’s Hacktron Partnership

· transkrypcja
NVIDIA and Korea’s Ministry of Science and ICT are expanding work on sovereign AI infrastructure, NVLink Fusion and Korean models, while NVIDIA names investments involving Point2 Technology and plans for Motif Technologies, RLWRLD and Upstage. Also covered: Anthropic’s reported gas-fired power deal…

246M+ Commits a Day. Your Complete Guide to Securing AI Factories | Aaron Stanley | S3 E6

GitHub says monthly commits on its platform went from 1.4 billion in April to 7.38 billion by September 2026. Some engineering teams now run software factories, where agents write, test, and ship code and no human reviews the diff. Most AppSec programs were built for a person writing code, another …

S1E09. Securing Agents

· transkrypcja
An agent can be turned against the person it works for when untrusted text, private data, and a way to act line up together. Simon Willison calls that the lethal trifecta. This episode sets that frame next to the OWASP list for 2026, then three public cases: a 2024 report on Slack AI, a 2025 Micros…

CodeRabbit’s Codex Output, OpenAI’s Control Room, and Meta’s Muse Ember

· transkrypcja
CodeRabbit CEO Harjot Gill discusses Codex-assisted software delivery and CodeRabbit Secure, while OpenAI details its Control Room for stopping research workloads and Meta introduces Muse Ember for repetitive agent tasks. Also covered: DeidSharp’s clinical-text de-identification evaluation, propose…

Tell the model the code is its own, and it stops repeating itself

Google DeepMind's researchers took the Signal protocol, what Wired calls a widely used encryption protocol, and proved it works as intended. Not tested it. Proved it. Proved it how? Formal verification. You write down mathematically what a protocol is supposed to guarantee, then you produce a proof…

Microsoft Hit by Anthropic's Claude Meetings, Sheets and Data

AI news today: Anthropic launches Claude Meetings, Claude Sheets and Claude Data as Microsoft shares close down 2.2%, OpenAI opens a Business Hub beta inside ChatGPT, and the AI Futures Project moves up its forecast for AI that matches the best human AI researchers. The conversation opens with Ant…