The AI Friday Brief
Good Saturday, NOLA. Today's vibe: The AI regulation story just got real. OpenAI's GPT-5.6 arrived hours after the White House asked them to slow roll it, Anthropic's Mythos is back in action for over 100 companies, and the entire frontier AI landscape is now operating under government vetting. Plus: what happens when 2,000 people try to hack your AI assistant, and why open-source models are becoming the only viable path for most of the world.
The Government Vetting Era Begins
GPT-5.6 is here, but under government oversight
OpenAI just launched GPT-5.6 in a tiered rollout — Sol (flagship), Terra (balanced), and Luna (fast) — less than 24 hours after the White House reportedly asked them to stagger the release. The company says this shouldn't become the default, but it's clearly the new reality. Deep dive on what this tier system means for access.
OpenAI blog / TechCrunch
Anthropic's Mythos is back — for approved organizations
After a two-week standoff with the Trump administration, Mythos 5 is live again for over 100 US companies and agencies. This settles the biggest AI policy drama of the week, but it confirms what we're seeing with OpenAI: frontier models now require explicit government approval to distribute.
Semafor / The Verge
The Reality Check: What Actually Works
What happens when 2,000 people try to hack your AI assistant
Fernando Irarrázaval ran an open challenge to break his OpenClaw test assistant, and the results are wild — people tried prompt injection, jailbreaks, token smuggling, and more. Spoiler: most attacks failed, but some succeeded in hilariously unexpected ways. It's a masterclass in how real adversarial testing actually works (and what it tells you about model robustness).
Hacker News
Open-source models are becoming the only viable path for most developers
The gap between open-weights and closed-source models is narrowing, but the real story is cost and access. As frontier models require government vetting and tier-based rollouts, open-source becomes the practical choice for building in resource-constrained regions and for teams that can't wait for approval. The regulatory environment is accidentally accelerating open-source adoption.
Hacker News / Doubleword
Smart model routing: pick the right AI for each task, in Cursor and Claude
This tool lets you define which model handles which part of your workflow — so you're not overpaying GPT-4 prices for tasks a smaller model can handle. It integrates directly into Cursor and Claude's code editor. Simple idea, massive cost savings.
Hacker News / GitHub
Worth a Read
The hidden labor draining AI productivity gains
Workers are saving time on individual tasks, but spending hours "botsitting" — feeding context, verifying outputs, fixing hallucinations, cleaning up messes. This podcast explores why the aggregate time savings aren't showing up in productivity data.
AI Daily Brief podcast
It's not about Anthropic vs. OpenAI anymore
The competition has shifted. AI capabilities are now advanced enough to have real political consequences, which means building these models is no longer just a tech problem — it's a policy and governance problem. This changes everything about how companies plan rollouts and who their stakeholders are.
TechCrunch / Hacker News
Why big labs are now hiring philosophers
As we saw earlier in the week: AI labs are wrestling with questions that engineers can't solve alone. Hiring philosophers isn't whimsy — it's a practical response to alignment, governance, and safety problems that are fundamentally philosophical.
The Economist
Quick Takes
DeepSeek open-sources 60–85% faster inference optimizations
DeepSeek released inference speed improvements that significantly reduce latency without sacrificing accuracy. If you're running any open-source model, this is worth benchmarking.
Hacker News / GitHub
NYT alleges Microsoft built a copyright-infringing supercomputer for OpenAI
The copyright lawsuit against OpenAI and Microsoft just took a new turn. After the Supreme Court ruled against Sony in a related case, NYT is repositioning its claims to argue that Microsoft's infrastructure investment was specifically designed to facilitate copyright infringement.
Ars Technica
Also
- AI children's books, body horror edition — Funny, unsettling, and surprisingly insightful look at what happens when you ask AI to generate kids' book illustrations.
- AI in mathematics is forcing big philosophical questions — As AI tools prove theorems and discover proofs, mathematicians are rethinking what it means to understand mathematics.
- The gap between open weights and closed models is narrowing — Technical breakdown of where open-source stands relative to frontier models — and why the regulatory environment matters.
- Wispr Flow: speak naturally in any app, get polished text output — Voice-to-polished-text for any app. Useful if you prefer dictating but need send-ready prose.
- Incident Report: CVE-2026-LGTM — A spectacular hypothetical security incident: two AI code review agents from competing vendors, both attached to a downed infrastructure project, making it worse.
- OpenAI poaches Uber India chief to lead expansion in its biggest non-US market — Strategic move: OpenAI is serious about international presence and hired someone who knows how to operate in complex regulatory environments.
- Why frontier AI labs are all building their own chips — OpenAI's Jalapeño, Meta's custom silicon, others following — the era of total Nvidia dependence is ending.
Pass it on