Episode Details

Today's topics:

GPT helps crack Erdős conjecture - An amateur used GPT-5.4 Pro to spark a novel proof idea for an Erdős conjecture on primitive sets; experts say humans still had to verify and rewrite the argument. Keywords: Erdős, primitive sets, GPT-5.4, proof verification, Terence Tao.

Talent and compute arms race - Thinking Machines Lab and Meta are trading researchers while big cloud deals unlock scarce Nvidia compute; the pattern repeats with Anthropic funding and Meta’s AWS CPU expansion. Keywords: talent mobility, GB300, cloud commitments, infrastructure access, valuation.

Cloud security risks in 2026 - Wiz’s 2026 retrospective says most breaches still come from misconfigurations, exposed secrets, and known vulns—AI mainly expands the surface area and speeds attacker workflows. Keywords: cloud misconfig, supply chain, identities, integrations, AI reconnaissance.

AI agents: memory and repo crawling - Anthropic is pushing persistent Memory for managed agents, while Claude Code experiments like “Bugcrawl” hint at full-repo scanning—raising stakes for governance and token budgets. Keywords: agent memory, audit logs, Claude API, repo analysis, enterprise controls.

Evaluating and measuring AI coding - Teams are building evaluation stacks because LLM testing isn’t deterministic, while debates grow over inflated “AI wrote X% of code” dashboards that can mislead leadership. Keywords: LLM eval, CI regression, LLM-as-judge, attribution metrics, ROI.

Distributed training across data centers - DeepMind’s Decoupled DiLoCo trains across regions with looser synchronization, aiming to keep runs going through outages and reduce networking bottlenecks. Keywords: distributed training, resiliency, WAN bandwidth, self-healing, scaling.

Generative vision models do perception - The “Vision Banana” paper claims an image generator can be tuned into a strong general vision system by expressing tasks as image outputs, blurring the line between understanding and generation. Keywords: generative pretraining, segmentation, depth, unified vision, benchmarks.

Sovereign AI: hype versus reality - A critique argues most enterprises don’t need nationally branded frontier models; the real requirement is sovereign deployment—data residency, auditability, and control of data flows. Keywords: sovereign AI, data residency, open models, vendor lock-in, compliance.

AI product trust, pricing, and UX - Google may move Gemini toward credits, Canva fixed a politically sensitive text-alteration bug, and OpenAI published new principles—together highlighting trust, pricing, and governance pressures. Keywords: usage credits, safety, transparency, content integrity, policy.

Real-world AI: the agent-run store - A San Francisco shop run by an AI agent made bizarre inventory choices and lost mo