Key Highlights
- Anthropic goes on the record about recursive self-improvement — in an essay dissected at length by Theo (t3.gg), Anthropic reports its engineers now ship ~8x as much code per quarter as in 2021–2025, with >80% of merged code authored by Claude, and openly floats the idea of a verifiable, coordinated pause on frontier AI if rivals would do the same.
- OpenAI files confidentially for an IPO, a week after Anthropic’s filing — even as the U.S. government reportedly negotiates a 1–5% equity stake in OpenAI to route into a public wealth fund. Both stories sharpen the economics-and-regulation question hanging over the labs.
- Apple’s WWDC26 Siri AI overhaul lands, built on Apple models developed alongside Google’s Gemini — and the theme rippled across sources (Gemini in Xcode, Private Cloud Compute extending to Google Cloud, and analysts reassessing Apple’s “slow and steady” bet).
- Google ships an agentic NotebookLM (now on Gemini 3.5 + Antigravity, with a code-executing cloud computer and PDF/Excel/PPT output) and opens Gemini to Apple developers via the Foundation Models framework.
- NVIDIA’s NVFP4 recipe delivers 4-bit pretraining on Blackwell at 1.31x–1.73x over FP8 with no measurable accuracy loss.
Analysis & Opinion
OpenAI files confidentially for IPO, following Anthropic — TechCrunch
OpenAI submitted a confidential S-1 to the SEC, just over a week after rival Anthropic did the same, intensifying the race between the two labs to tap public markets. The company — valued at $852 billion in its latest round — said timing is undecided and “may be a while,” and unusually paired the filing with a philosophical statement on its AGI mission during what is normally a regulatory quiet period. The financial backdrop is heavy: OpenAI reportedly missed user and revenue targets, projects burning $85 billion in 2028, and expects to spend roughly its entire $122 billion recent round on compute, putting profitability years out. 2026 is shaping up as a landmark year for tech listings, with OpenAI, Anthropic, and a ~$1.75 trillion SpaceX all potentially debuting.
Washington wants a piece of OpenAI — Rundown
The U.S. government and OpenAI are reportedly negotiating a deal that would hand Washington a 1–5% equity stake, directed into a public wealth fund so ordinary Americans could share in AI gains. Sam Altman has met with Senator Bernie Sanders and Trump administration officials, and the proposal aligns with OpenAI’s April industrial-strategy paper; President Trump called it “a beautiful thing.” Critics, including former AI czar David Sacks, warn it accelerates “corporate-government fusion,” raising conflict-of-interest concerns when the state simultaneously owns, profits from, and regulates the same firm. The open question is whether any wealth-sharing reaches ordinary citizens or stays theoretical — a live test of how the AI windfall gets distributed.
Our latest fraud and scams advisory — Google
Google’s Trust & Safety teams peg global fraud losses at nearly $580 billion for 2025, with roughly one in five adults victimized. The advisory flags four escalating threats: phishing that defeats MFA via Adversary-in-the-Middle and QR-code attacks; crypto- and AI-fueled investment scams (Americans lost over $11 billion to crypto schemes in 2025 per cited FBI figures); malicious finance apps that request excessive permissions then deploy extortion malware after install to dodge store review; and impersonation of law enforcement across South/Southeast Asia and Gulf regions using coordinated email and video-call campaigns. The throughline is attackers weaponizing AI and trusted cloud services to scale social engineering — a direct counterpoint to the productivity-gain narrative dominating the rest of the day’s news.
Why Apple’s slow-and-steady AI bet is starting to look pretty smart — TechCrunch
After years of criticism for lagging, Apple is being reassessed as its deliberate approach — culminating in a Gemini-powered Siri AI — looks less like falling behind and more like avoiding the industry’s missteps. Craig Federighi jabbed at rivals “racing forward… pursuing AI for the sake of AI,” framing Apple’s caution as a response to consumer skepticism over job loss and AI harms. The catch: the new Siri capabilities won’t reach consumers until a later-2026 beta.
New Products & Tools
Apple’s new Siri AI overhaul is here (sort of) — Rundown
At WWDC26 Apple unveiled an upgraded, rebranded Siri AI running on Apple’s own models built alongside Google’s Gemini, able to read on-screen context, pull from apps, and act across the system, with a dedicated chatbot app and private cross-device history. It ships free this fall for iPhone 15 Pro and newer (public beta next month, excluding EU and China), though observers noted the demos felt closer to “2024-level” models than the current frontier.
Bringing the latest Gemini models to Apple developers — Google
Starting with iOS 27, developers can call cloud-hosted Gemini through Apple’s Foundation Models framework (via the Firebase Apple SDK), using one API to swap between on-device Apple models and cloud Gemini. Xcode also gains an agentic Gemini experience for reviewing code, fixing bugs, and building features.
Do better research with NotebookLM — Google
NotebookLM gained agentic chat and deeper reasoning on Gemini 3.5 + Antigravity, backed by a secure cloud computer that runs code across 100+ skills and a >65% average win rate over the prior version. It now outputs PDFs, spreadsheets, slides, and visualizations, and can run its own Google searches to assemble sources.
Expanding Private Cloud Compute — Apple Security
Apple is extending Private Cloud Compute beyond its own data centers for the first time, running Apple Intelligence workloads on Google Cloud with NVIDIA Confidential Computing, Intel TDX CPUs, and Google’s Titan chip. It keeps PCC’s core guarantees — stateless compute, non-targetability, verifiable transparency — and adds a cryptographically verifiable, append-only ledger of all Google Cloud PCC hardware to counter supply-chain threats, with binaries to be published for Security Bounty research.
Powering the future of robotics in Europe — Google
Google DeepMind launched a three-month accelerator for 15 early-stage European robotics startups (Norway to Switzerland) spanning healthcare, manufacturing, waste, and climate, offering access to its models, mentorship, and partner network.
How an e-scooter founder raised $5 million to build space data centers — TechCrunch
Spin founder Euwyn Poon raised a $5M seed (a16z Speedrun and others) for Orbital, which aims to put AI inference compute in space — eventually 10,000 satellites at ~1 gigawatt — betting on cheaper Starship launches, with a demo flight carrying an NVIDIA chip and first data-processing craft targeted for 2028.
Research
Train Models Faster with JAX and MaxText Using NVFP4 on NVIDIA Blackwell — NVIDIA
NVIDIA’s NVFP4 recipe in TransformerEngine enables 4-bit mixed-precision JAX pretraining on Blackwell with no measurable accuracy loss versus FP8, using two-level microscaling, selective Random Hadamard Transforms, and stochastic rounding to hit 1.31x–1.73x speedups while preserving convergence.
Interviews & Conversations
I didn’t expect this from Anthropic — Theo - t3.gg (44:57)
In this transcript-based summary, Theo walks through Anthropic’s essay arguing that AI is already measurably accelerating AI development — engineers now ship ~8x the code per quarter versus 2021–2025, >80% of merged code is Claude-authored, and a poll of 130 research-team staff estimated ~4x output with the “Mythos” preview. He highlights the safety core: the length of tasks models can complete autonomously is doubling every ~4 months (down from 7), an automated Claude reviewer would have caught roughly a third of past production incidents, and Claude recovered 97% of a research gap two humans recovered only 23% of — but he stresses these are 50% success-rate figures and that “research taste” remains a human edge. Most striking, Theo flags Anthropic’s three scenarios (stall-and-diffuse, compounding-efficiency, full recursive self-improvement) and its explicit willingness to slow or temporarily pause frontier development if rivals verifiably did the same — a nuclear-treaty-style proposal Anthropic admits is far harder to verify than missile silos. He ties it to unsettling alignment research (subliminal “owl-loving” trait transfer via seemingly random numbers; emergent misalignment generalizing from one domain to all), warning that self-training models could compound misalignment in ways humans can’t inspect — landing on “build a safety net, not guardrails.”
References
- OpenAI files confidentially for IPO, following Anthropic — TechCrunch, 2026-06-08 [blog]
- Washington wants a piece of OpenAI — Rundown, 2026-06-08 [blog]
- Our latest fraud and scams advisory — Google, 2026-06-08 [blog]
- Why Apple’s slow-and-steady AI bet is starting to look pretty smart — TechCrunch, 2026-06-09 [blog]
- Apple’s new Siri AI overhaul is here (sort of) — Rundown, 2026-06-09 [blog]
- Bringing the latest Gemini models to Apple developers — Google, 2026-06-08 [blog]
- Do better research with NotebookLM — Google, 2026-06-08 [blog]
- Expanding Private Cloud Compute — Apple Security, 2026-06-08 [blog]
- Powering the future of robotics in Europe — Google, 2026-06-09 [blog]
- How an e-scooter founder raised $5 million to build space data centers — TechCrunch, 2026-06-09 [blog]
- Train Models Faster with JAX and MaxText Using NVFP4 on NVIDIA Blackwell — NVIDIA, 2026-06-08 [blog]
- I didn’t expect this from Anthropic — Theo - t3.gg, 2026-06-08 [video]