The Trust Tax

• AI makes shipping features fast. It doesn't make them good.
• Linear measures quality by edge cases missed and manual corrections made, not features shipped.
• Taste is a trainable skill, not a personality trait.
• The quality wedge opens the moment you can generate code in seconds.
• The agent generates. The human validates. The work shifts from creation to curation.

We have automated techniques that can cover, as far as we can tell, the full space of vulnerability-inducing bugs.

• Firefox 150 includes 271 vulnerabilities found and patched using Mythos Preview.
• Bobby Holley says AI now covers the full space of vulnerability-inducing bugs.
• The 'bootcamp' metaphor frames this as a finite, mandatory overhaul for all software.
• Open source maintainers lack the bandwidth and access to replicate Mozilla's approach.
• Raffi Krikorian warns that resource inequality will perpetuate security gaps across the ecosystem.

This was understandably confusing for the 98% of folks not part of the experiment, and we've reverted both the landing page and docs changes.

• The pricing page update was a fake-door test, not a shipped change.
• They were testing whether to roll best models to all plans for Codex users.
• The page and docs were reverted within hours after community pushback.
• The test created confusion rather than clarity, undermining trust in the product roadmap.

there is little to no difference between pi and codex when it comes to the loop. almost all of the below has nothing to do with the harness.

• The agent loop design is converging across major platforms.
• Context compaction, not loop architecture, is the actual bottleneck.
• Pi uses a 20k-token recency window excluded from summary; Codex keeps all turns.
• Both approaches prioritize tool state preservation over raw context retention.
• The differences are engineering trade-offs, not architectural moats.

• Google is shipping separate TPU-8t (training) and TPU-8i (inference) chips.
• The 8i is optimized for streaming token generation with lower per-token latency.
• Agentic workloads generate more inference tokens per user than any previous application.
• Compute cost scales with session length and context window, not just model size.
• Open source inference frameworks are racing to match proprietary inference stacks.

Wrote up Anthropic's self-own about Claude Code pricing from this afternoon on my blog - it turned out they'd reversed course just as I hit publish

• Anthropic tested removing Claude Code from $20 Pro tier
• Only 2% of new signups were affected by the test
• The pricing page was accidentally updated for everyone
• They reverted the page but kept the test running

Every piece of software is going to have to make this transition, because every piece of software has a lot of bugs buried underneath the surface that are now discoverable

• Mozilla found 271 vulnerabilities using Mythos Preview
• AI tools can now cover the full space of vulnerability-inducing bugs
• Open source projects face particular challenges with resource constraints
• Engineering leaders pulling thousands of engineers to work on this for six months
• This represents a finite but difficult transition moment

tongiht i have had a couple of drinks

• Sam Altman admits to drunk tweeting
• Followed by offering rate limit resets
• Part of a series positioning OpenAI as the generous alternative

• TPU 8t for massive model training, TPU 8i for low-latency inference
• Single superpod scales to 9,600 chips with 2 petabytes shared memory
• 121 ExaFlops of compute with near-linear scaling to million chips
• Targets over 97% goodput through comprehensive RAS capabilities
• TPU 8i pairs 288 GB high-bandwidth memory with 384 MB on-chip SRAM

• Lumma Stealer infection at Context.ai led to OAuth token theft
• Attack chain: malware → OAuth tokens → Workspace account → Vercel internals
• Environment variables not marked sensitive were readable with internal access
• Two-month dwell time from infection to disclosure
• CEO attributed attacker velocity to AI augmentation

As expected. Claw issue, not harness issue.

• OpenAI's Codex OAuth issues attributed to implementation not harness design
• Pi harness users experiencing similar OAuth failures
• Fallback path to pi when Codex OAuth fails is also broken