
Sign up to save your podcasts
Or


Anthropic’s Claude Duet brings persistent, background AI assistance to connected apps—but its approval gates and unseen browser raise questions about how well agents can be monitored. This roundup also covers METR’s evaluation concerns, Recursive’s local coding model Flint, Unitree’s security humanoid, Google and Meta’s TorchTPU, OpenAI’s Orchestra Bio acquisition, and the financing pressures behind the AI infrastructure boom.
Beacon’s launch puts agent memory, telemetry, and real-time security controls in focus, alongside questions about shrinking entry-level technical hiring and Databricks’ autonomous marketing agents. We also unpack Stanford’s parallel text-generation research, compare three open-weight models on Nebius Token Factory, and examine why reproducible lab data may matter more than model architecture in biotech.
DeepSeek V4.1 Flash, Meta’s Muse Spark Open, Tencent’s Hy Memory 2.0, and NVIDIA’s Nemotron 3.5 all point toward a new priority in open AI: getting more useful work from fewer tokens and less compute. We unpack their capabilities, costs, benchmarks, licensing and caveats, plus Kaggle’s new model-identification competition and key discount deadlines.
Agents that work around web defenses and read reports about themselves, a covert recording device discovered at OpenAI, and Character.ai’s Chapter 11 filing lead this roundup of the past 48 hours. We also cover Google’s Harness Squared research, new tools for improving training data and robotics, fresh model launches, and a forecast that ties AGI to GPU prices.
Explore the rapid shift toward hyper-specialized AI as Z.ai undercuts major frontier models on legal workflow costs, while Meta debuts Muse Wonder and Microsoft unveils its Maia 300 chip.
We also break down new safety disclosures from Anthropic, escalating researcher disputes at OpenAI, and breakthroughs allowing massive models to run on consumer hardware.
James Turner breaks down the race for low-latency AI, exploring OpenAI's new Ultrafast tier and the rise of non-autoregressive decision models from Liquid AI and Microsoft. Plus, a look at Google's universal Gemini agent, emerging agent protocols, and Anthropic's latest defense initiatives amid ongoing legal and financial headwinds.
James Turner breaks down OpenAI's launch of GPT-6 featuring dynamic Intelligent UI, Anthropic's aggressive pricing moves with Claude Haiku 5.5, and Mistral's trillion-parameter release. Plus, explore the implications of 722 AI-generated, machine-verified math research papers dropped on GitHub.
James Turner explores OpenAI's new invisible watermark textGrain, major open-source MoE releases from Reflection AI and Aleph Alpha, and ScienceLogic's surgical controls for enterprise agent swarms. Plus, massive funding updates for chipmaker Etched and Anthropic's new enterprise training initiative.
Open source decision models are landing fast, with Kev and Clef offering drop-in alternatives that let developers swap endpoints and models without rewriting apps. We also dig into why agents need more than a kill switch, from human escalation and swarms to the case for snapshots, logs, and deterministic rollback.
This episode covers NVIDIA’s Open Agent Safety Platform, including the Sentry out-of-band watchdog and OpenShell secure runtime boundary, plus Google DeepMind’s Gemini 4 Argon and its one million token output window for defenders.
Also discussed: OpenAI’s Dots and new computer-use stack, the LIFT research paper on latent-state feedback transformers, and quick hits on robot world action models, Solar Mini 4, and OpenAI’s Jalapeño inference chip.
From the publisher's feed