This Week in AI

O'Reilly Live · June 15, 2026

Host

YK Sugi

Guest

John Lindquist

Quick picks

AI's self-improvement?

Fast Gemma Challenge (Google + Hugging Face): coding agents collaborate to serve Gemma 4 E4B faster - speeding up inference (tokens/sec) without degrading quality. The model is unchanged; the agents do the optimization. post · dashboard
Microsoft SkillOpt (open source, MIT): instead of fine-tuning the model, it keeps improving an agent's plain-text instructions file - like training, but the "weights" are words. ~23-25% accuracy gains; ships as a small text file. VentureBeat · GitHub
Anthropic, "When AI builds itself": Claude already authors 80%+ of code merged at Anthropic, and its typical engineer now merges ~8x as much code per day as in 2024. It's also increasingly doing the research - picking the better next research step 64% of the time (up from 51%). essay

1

Launched June 9 - then disabled June 12

Cybersecurity and biology/chemistry questions fall back to Opus 4.8
On AI-training questions, the model is designed to perform worse - no visible fallback
The goal: keep competitors from using it to build their own models
system card, p. 12

June 12 - US export control directive

Directive bars foreign nationals (inside and outside the US) from Fable 5 and Mythos 5; Anthropic disabled both models for all customers rather than selectively block
Amazon reportedly got Fable 5 to surface info useful for cyberattacks; its CEO raised it with senior officials, and Commerce Secretary Lutnick sent the order to Anthropic
The government's position: the order came after Anthropic refused to patch or de-deploy. Senior Anthropic staff went to Washington DC to meet White House officials
CNBC · Axios · NBC News

Anthropic: minor, previously-known vulnerabilities also present in other public models (names OpenAI's GPT-5.5); a narrow jailbreak shouldn't pull a model used by hundreds of millions statement
But neither side has published the technique or the specific flaws - the "other models too" claim isn't a demonstrated head-to-head
Independent researcher Katie Moussouris (CEO, Luta Security) says she's seen the report: "it's not a jailbreak" - defensive prompting that found known vulnerabilities - and calls it "a complete overreaction" / "an own goal" her post

2

The agentic coding shift

Boris Cherny, Claude Code creator, at the Sequoia AI Ascent talk: Claude writes 100% of his code
Dozens of PRs a day, mostly from his phone, hundreds of agents running via /loop
"We have no more manually written code anywhere at the company"
Caveat: big, complex codebases still struggle - "wait for the next model"
Another perspective, across AI tools (not just Claude): 43% of AI-generated code changes need debugging in production VentureBeat

3

Full-year 2026 AI coding budget gone by April (reported late May)

Spent mostly on Claude Code and Cursor
84-95% of ~5,000 engineers use AI tools monthly
COO: can't yet link the spend to "25% more useful consumer features" - "that link is not there yet"
Response: $1,500/month cap per employee
Fortune · Inc

4

Company	What	Valuation	Source
SpaceX	IPO June 12 (Nasdaq: SPCX); absorbed xAI (~$250B) in an all-stock deal in February	$1.75T at IPO (largest ever); raised ~$75B	CNBC
OpenAI	Filed S-1 on June 8	$852B (last private round)	NBC News
Anthropic	Filed S-1 on June 1, targeting Oct; now valued above OpenAI	$965B (Series H, May 28)	TechCrunch
Google	$84.75B equity + $85B+ debt (incl. a rare 100-year bond)	~$4.4T market cap (public)	CNBC · CNN