GPT-5.5 ships, Copilot's $14K-of-usage party ends

The release pairs a capability bump with a personalization push: ChatGPT can now use memory sources and connected Gmail to tailor answers, with Instant getting noticeably better at remembering preferences and matching tone.^{[2]OpenAI — GPT-5.5 Instant} The "smarter" claim is anchored in factuality work — health-related HealthBench scores rose between +1.8 and +5.5 points across subsets — but the system card is candid about regressions: small but statistically significant drops in gore (0.703) and sexual-content (0.806) filter pass rates relative to GPT-5.3 Instant, which OpenAI says it addresses with system-level mitigations rather than further fine-tuning.^{[3]OpenAI — GPT-5.5 Instant System Card}

The Preparedness flag matters because it's the first time an Instant-tier (i.e., always-on, non-reasoning) model has crossed High Capability in two domains simultaneously. OpenAI is now shipping that capability to every free user by default.^{[3]OpenAI — GPT-5.5 Instant System Card}

MRC is what runs Stargate, the OCI-hosted training clusters, and Microsoft's Fairwater deployment. The technical pitch: at 131,000+ GPUs the bisection-bandwidth math breaks down with conventional Clos topologies, so OpenAI built a multi-plane fabric with hardware-assisted load spreading that keeps utilization high even under hotspot-heavy collective ops.^{[4]OpenAI — Supercomputer networking}

The strategic move is the open-sourcing. OpenAI is positioning itself as the reference implementation for hyperscaler-grade AI fabrics — co-publishing with the silicon vendors who actually build the gear, rather than locking the design behind partner NDAs. That maps cleanly to OpenAI's broader pattern this year of publishing its harness, orchestration, and now networking primitives to seed an ecosystem rather than monetize the layer directly.^{[4]OpenAI — Supercomputer networking}

OpenAI is publishing three principles around the platform: answer independence (ad placements never alter the substantive answer), conversation privacy (advertisers don't see chat content), and user control.^{[5]OpenAI — New ways to buy ChatGPT ads} These are likely the lines OpenAI will be held to as ad load scales — the principles are easy to assert at low load and harder to keep at the spend levels CPC implies.

The shift to CPC matters more than the Ads Manager UI: it tells advertisers ChatGPT Ads is willing to be measured against direct-response benchmarks (clicks, conversions) instead of just brand-impression metrics, which is what serious budget allocation requires.^{[5]OpenAI — New ways to buy ChatGPT ads}

Two adjacent shipments make this stickier: Microsoft 365 add-ins (Excel, PowerPoint, Word) are now generally available with Outlook coming, and the data-connector list expanded by eight (Dun & Bradstreet, Fiscal AI, FMP, Guidepoint, IBISWorld, SS&C IntraLinks, Third Bridge, Verisk), plus a Moody's MCP app covering 600M+ companies.^{[1]Anthropic — Agents for financial services} The play is direct: instead of selling Claude as a horizontal LLM, Anthropic is shipping a vertically integrated stack for the buyer who values domain coverage (Moody's, Verisk) and Office-native distribution.

The 100% Walleye number is the marquee customer claim and is the one to watch for replication — full-fund adoption is hard to buy and harder to sustain.^{[1]Anthropic — Agents for financial services}

Gemini API File Search — multimodal RAG

Three changes ship at once: multimodal indexing (text + images + PDFs + scientific imagery), per-document KV metadata you can filter on at query time (so you can scope a search to a customer / region / status), and page-level citations so you can verify where an answer came from. The 50%+ context-window reclamation isn't a quality claim, it's a cost claim: precise retrieval lets you avoid stuffing whole documents.^{[6]Google — Gemini API File Search multimodal}

Gemma 4 MTP drafters

The drafter models ride speculative decoding: they share the target model's KV cache and reuse activations on the edge variants (E2B, E4B). Google reports ~2.2× speedup on Apple Silicon at batch 4–8 and similar gains on Nvidia A100. The drafters are Apache 2.0, compatible with vLLM, SGLang, MLX, Ollama, Hugging Face Transformers, and LiteRT-LM — i.e., usable in every serving stack without re-licensing.^{[7]Google — Accelerating Gemma 4}

~00:00 Opening framing: "AI reinvented the computer industry" — Jensen's five-layer cake (energy, chips, infrastructure, models, applications) and why this is a re-industrialization, not a software cycle.

~04:00 The compute math: agentic AI ~1,000× more compute per task, ~100× more users than generative AI. This is the core bull case for the entire CapEx wave.^{[17]Milken Institute — Jensen Huang}

~12:00 Hardware: Vera Rubin racks at $4–5M each, 3 tons, 1.5M parts, 7 distinct chip types, football-field-scale assemblies. The cost-per-rack is the real barrier to entry, not the chips.

~19:00 Re-industrialization: half a trillion in orders to bring suppliers to the US; "several trillion dollars" of opportunity ahead.

~25:00 Investment posture: anchor-model checks for CoreWeave, Nebius, NScale (~$1 Nvidia attracts ~$9 from others). OpenAI investment "likely the last" because AI-native model providers became strongly gross-margin-positive in the last 3–6 months.

~31:00 China and export controls: backs H200 exports, opposes denial. "100% of you need AI. None of you should have a nuke."^{[17]Milken Institute — Jensen Huang}

~36:00 Cyber: defense will look like "swarms of white blood cells" — open-source agents patrolling for asymmetric defense rather than parity with attackers.

~40:00 Doomer rebuttal: explicitly takes on Hinton's 20–30% extinction-risk figure; uses radiology-jobs counterexample to push back on labor-collapse claims.

~44:00 Defense posture: "CEOs aren't elected, shouldn't be the ones gating US military's use of AI."

~48:00 Closer: AI compresses research months into days. "100x your ambition."

"100% of you need AI. None of you should have a nuke."

"100x your ambition."

~00:00 Intro and Lupsasca's bona fides — what makes this guest different from the usual "AI for science" guest is that he's already published with the model as a co-author, not just used it as a tool.

~10:00 The first paper: GPT-5.2 Pro conjectured the closed-form gluon tree amplitudes, internal OpenAI model spent 12 hours and produced the proof. Lupsasca describes the workflow as "vibe physics" — driving the model from a handful of physics intuitions and verifying each derivation.

~25:00 The graviton sequel three weeks later, done end-to-end by the public GPT-5.2 Pro — the gap between "internal OpenAI model" and "what's in the API" is closing fast.

~40:00 Day-to-day craft: parallel "scout" chats down different solution paths, using GPT to reduce time-stuck-confused. The bottleneck is no longer the math, it's deciding which math to do.

~60:00 Hot takes — arXiv slop, raising the bar to "community-decades-stuck" problems, why static papers are obsolete and Lean / formal-methods verification is making a comeback as the model's bottleneck.

~75:00 Taste and question selection as the remaining human edge — and why this changes who gets to be a competitive theorist.

~85:00 Closing: where Lupsasca thinks "vibe physics" hits a wall (data-anchored fields with physical experiments) and where it doesn't (closed-form math).

~00:00 Why small-model serving is different — not a "smaller LLM," fundamentally a different traffic shape (more tokens-per-second per host, batching that benefits from different scheduling).

~04:00 Sie architecture and what it deliberately omits from vLLM-style designs.

~08:00 KEDA-driven autoscaling on real production traffic with Prometheus metrics — the operational story behind the engine.

~12:00 MTEB benchmark numbers for the supported model families (BERT, Qwen, ColBERT).

~15:00 Multi-tenancy in production — what changed when traffic spiked.

~18:00 Closing call to action and where Sie fits vs Triton/TGI/vLLM.

~00:00 Why Google is investing this hard in edge — privacy, latency, and the cost equation when CPM-priced LLM calls hit mobile traffic shape.

~05:00 LiteRT-LM stack overview — what changed from TensorFlow Lite era.

~12:00 NPU acceleration: 13× speedup over CPU, calibration-aware quantization story.

~18:00 iOS demo at ~56 tok/s — the live benchmark moment.

~23:00 35× vs llama.cpp comparison — caveats, model sizes, and what's actually being measured.

~29:00 Multi-token-prediction drafters tie-in (cross-references the Gemma 4 MTP launch from the same day).

~34:00 What's next: the Gemma 4 31B-on-laptop story.

~38:00 Q&A.

~00:00 Raj introduces himself (staff SWE on IKEA's deliverance-services domain) and frames the problem: KBs go stale, tribal knowledge stays in Slack threads.

~05:00 The flip: instead of capturing knowledge proactively, run agents on real tasks and let them fail — failures become a queue of missing-context tickets.

~12:00 The KB shape — what gets indexed, the per-document metadata schema, and how it links back to the failing tool calls.

~23:00 Demo: real IKEA workflow where the agent failed on a delivery-routing question, the failure surfaced an undocumented business rule, and the rule went back into the KB.

~33:00 Versus traditional RAG: why eager retrieval over a giant corpus loses to lazy retrieval over a living, failure-curated set.

~45:00 Operational mechanics: who owns the queue, how tribal-knowledge owners get paged, and why the loop has to close in days, not quarters.

~55:00 Q&A on adversarial cases (what if the agent fails for the wrong reason?).

~00:00 Where Theo and Prime fully agree: the easy-subsidy era is over, Cursor's pricing was the canary (July 2024 switch from message-count), Anthropic's painted-door test on Claude Code's $20 tier was a real signal, and Dario's per-model profit framing ("each model is its own company") is correct.

~08:00 The pivot: Prime says it's about clawing back revenue. Theo says it's compute. Pausing Copilot signups is the proof point — you don't pause signups to extract more money from existing customers, you pause when you can't serve more.

~15:00 On Opus 4.6/4.7: Prime claimed they're losing money on training. Theo: they're post-training iterations (RLHF / RLVR), which is far cheaper than pre-training. The unit economics aren't what Prime thinks.

~22:00 Cost of intelligence is dropping when measured in tokens-per-task, not price-per-token. Artificial Analysis Intelligence Index numbers cited: $500 for 5.5-low at the same task quality as $2,850 for 5.4-high.

~29:00 Subscription subsidization math: Anthropic's consumer terms exclude commercial use and Max plans don't carry BAA/HIPAA — i.e., the cheap plans aren't designed to support enterprise workloads, they're loss leaders.

~38:00 GitHub Copilot's 7.5× multiplier on GPT-5.5 and 15× on Opus reflects GPU allocation, not model cost — a UI for compute scarcity, not a margin grab.

~45:00 Cursor Composer 2 is fine-tuned Kimi K2.5 from Moonshot — a Chinese model — which Theo flags as the kind of integration that's only viable when compute is the constraint.

~52:00 Enterprise inference spend: Uber burned its yearly AI budget in 4 months — a real number that suggests the demand side is real, not hype.

~60:00 Prime is right about Anti-Gravity, OpenCode plugin bans, and the broader vendor-lock pattern; Theo just thinks the why is different.

"You don't pause signups because you want to make more money."

~00:05 The pricing math: 1,500 premium requests × ~$10 of agentic-task compute on average = ~$14K. Theo's example: $115 of usage at 0.8% of subscription.

~03:53 The June 1 change: switch from request-based to AI-credit billing. Pro = 1,000 credits ($10 worth), Pro Plus = 3,900 credits ($39 worth). Priced per token per model. Code completions remain unbilled.

~06:00 The hot take: GitHub priced the plan for chat-style usage and got blindsided by agentic workloads. Hence the paused Pro Plus signups (consistent with Theo's compute-scarcity thesis above) and the upcoming overhaul. Until June 1, this is a ~27-day window.

~00:00 The opening parade of AI Twitter posts (Aaron Levy, Brian Johnson, Sam Altman switching to polyphasic sleep to keep up with Codex) — the early Genai narrative was time-saving but the agentic era has people working until 3am.

~04:00 The "infinite backlog" as a counter to the lump-of-labor fallacy: leaders' job is to prioritize from an always-larger queue.

~06:00 AI assistants got people from 1× to 2–4×; agents "break the rules of time" by replicating the worker infinitely.

~07:00 Closest existing analogy: entrepreneurship — the kierkegaardian "dizziness of freedom."

~10:00 Tang Yan: "the work no longer drains you through typing. It drains you through judgment. More attention, more context switching, more verification, more decisions per hour."

~15:00 The new org chart: agent ops engineers, context librarians, eval engineers, coordination architects, information pipeline owners, experiment portfolio managers, entrepreneur coaches. Aaron Levy at Box already hiring "agent engineering" FTEs to wire agents into Salesforce/Workday/Box.

"Agents make every job a startup. Not everyone wants to work for a startup, however, and that's going to cause stress and challenge."

"Are you building pacing infrastructure or just rewarding whoever stays up latest?"

~02:00 Ezra Klein's anti-doomer argument: Jevons paradox applies to knowledge work — better tools create more demand, not less.

~07:00 SWE demand is accelerating, not collapsing. NLW pulls labor data showing the post-2023 layoff narrative isn't borne out by JOLTS.

~12:00 Largest entrepreneurship explosion in history? — formation rates, AI-native startup count.

~15:00 Anthropic ARR doubling every six weeks (cited number) — the demand side of the compute story.

~20:00 CapEx bubble reframed: the backlog of AI deployments outpaces even hyperscaler spend, so "bubble" framing requires acknowledging unmet demand first.

~25:00 Sam Altman's recent messaging pivot — less "AGI is here," more "this is a productivity tool" — and what NLW reads into it.

~00:00 The anticipation gap: consumer agents demo well but hit a reactive ceiling because consumer AI inherits ChatGPT's query-box behavior — it waits for you. There is no consumer-AI equivalent of "Google ranks for this query without asking" yet.

~05:00 Why coding agents work but consumer doesn't — there's a compiler for code, no compiler for taste.

~12:00 Symphony and the OpenAI enterprise attention bottleneck — open-source orchestration alone doesn't fix the fact that someone still has to triage agent output.

~18:00 Fake proactivity vs real lived proactivity — the test isn't "does the agent message you," it's "does the message change a decision you'd otherwise make."

~23:00 Product roundup of the genuine attempts: Poke, Clickie.so, Cluely, Cowork + Stripe agent wallets, Codex's Chronicle memory feature.

~29:00 The trust ladder — Nate's five-step permissioning framework for agent autonomy.

~34:00 Prosumer bridge: how proactive agents will reach consumers via work first (Slack, Notion, Superhuman), then cross over.

~38:00 Three early warning signs proactive agents are coming: key hires (Peter Steinberger to OpenAI), monthly load-lifting tests, frontier release notes mentioning consumer agentic intent with memory.

~43:00 Specific cited datapoints: Hawaii swimsuit personalization, "lines in China to uninstall openclaw," GitHub planning for 10×–30× repo growth, his pushback on the ubiquitous "book a trip" demo.

~00:00 Setup: connecting Higgsfield to Claude via a custom MCP connector — Claude becomes the orchestration layer, Higgsfield is the model.

~04:00 Higgsfield Marketing Studio + Hypermotion style ads — what the actual output looks like, where it still breaks.

~10:00 Claude Code + Higgsfield CLI as a creative-studio project layout.

~15:00 Domain expertise as markdown — encoding "what works for our brand" as files Claude reads instead of as system prompts.

~20:00 gws CLI Google Sheets asset tracker — a real workflow piece, not a toy.

~25:00 Reference image consistency for product ads: same model + lighting across a campaign.

~30:00 Reverse-engineering skills from winning outputs — when something works, can Claude extract the prompt structure?

~35:00 Routines for scheduled autonomous ad generation — overnight runs with morning approval queues.

~40:00 Hot take: "this is the worst AI video will ever be" — accept the current ceiling, build for next year's.

~00:00 Video, writing, and research tools — claude-video, WRITING.md, paper2code.

~03:01 Skill quality and management — skill-doctor (local audit), skills-manager (Tauri, symlink sync across 27 platforms).

~03:01 Precision coding workflows — Matt Pocock's TDD primitives, Design Council's 11-agent debate system, Code Overhaul for architectural teardowns.

~03:01 AI-native design systems — Oh My Design (67 real-world systems), spec-first Web Design skill, Claude → SwiftUI transpiler.

~06:02 Library-specific skills — Library skills (embeds skills in NPM/Python packages), official Cloudflare skills, DSPy agent skills.

~05:02 Game dev / image generation — Sprite Pipeline, Agent Sprite Forge, OG Image, Shots (app-store screenshots via GPT Image 2).

~07:02 Token efficiency & session continuity — Usage Limit Reducer (visualizes JSONL token spend), Agent Session Resume (handoff checkpoints), 3M Framework (persistent background operator).

~07:02 Voice and writing style — Agent Style (Strunk & White), Aerodani Speak (Rocky from Project Hail Mary voice), Linus Torvalds skills.

~14:04 Multi-model review — God Mode (/work command sending code to Codex + Gemini + open model in parallel), Compose Performance skills.

datasette-referrer-policy 0.1

Datasette's default Referrer-Policy: no-referrer header was breaking OpenStreetMap tile loads (OSM blocks no-referrer requests). The plugin lets operators override per-site. Built with Codex + GPT-5.5.^{[8]datasette-referrer-policy}

datasette-llm 0.1a7

Configurable per-model defaults: pin a model + set temperature=0.5 for all enrichment operations, making Datasette LLM deployments reproducible and cost-predictable.^{[10]datasette-llm 0.1a7}

llm-echo 0.5a0

Adds -o thinking 1 support aligned with LLM 0.32a0's new thinking-block architecture — fully offline CI tests for reasoning workflows via uvx.^{[11]llm-echo 0.5a0}

The Andon Labs experiment is a clean reductio ad absurdum of the "let the agent run wild" pattern: a real cafe, real suppliers, real money — and an agent making decisions with no model of the physical kitchen it's allegedly operating. The 120-eggs anecdote isn't just a funny number; it's that a supplier received an order, allocated inventory, and the cafe owner had to send a follow-up email tagged EMERGENCY to undo it.^{[9]Simon Willison — Stockholm AI cafe}

Simon's framing is the more durable point: agentic systems should distinguish read-only reasoning from outbound actions that touch other parties. The supplier did not opt in to being part of the experiment; the EMERGENCY email confirms it.^{[9]Simon Willison — Stockholm AI cafe}

Cohen had built a 5% economic stake in eBay before the formal offer. Market reaction: eBay +5%, GameStop -5% to -10% on the day of the announcement. Michael Burry exited his full GME position.^{[14]Sherwood Snacks — GameStop eBay bid}

The strategic rationale Cohen pitched: GME stores as eBay live-streaming / authentication / fulfillment hubs, and slashing eBay's marketing budget. Morning Brew highlights Cohen's awkward CNBC interview where he couldn't explain the funding gap.^{[15]Morning Brew — GameStop offers eBay}

Analyst consensus: deal unlikely to close — eBay's existing turnaround is working, and the financing gap is too wide for a "highly confident" letter to plug without a strategic equity partner.^{[14]Sherwood Snacks — GameStop eBay bid}

The article ties the policy reversal to two things: David Sacks's late-March departure (he was the deregulation advocate) and the rising influence of more security-focused officials. Industry pushback is the predictable "this slows US competitiveness vs China" line.^{[13]Tech Brew — AI EO reversal}

Important caveat: the article notes discussions are still speculative — no final EO language exists yet. The voluntary early-access agreements with Google/Microsoft/xAI are real, the EO is not yet.^{[13]Tech Brew — AI EO reversal}

~00:00 Brockman frames the anecdote as the kind of capability that didn't exist 18 months ago — not just "GPT wrote some code," but an autonomous loop that profiled, identified bottlenecks, and iterated. The "overnight" framing matters: it's about wall-clock asynchrony, not raw speed. The implication for staffing math is the same one NLW raises in topic #13 — the constraint moves from typing to judgment.^{[23]Sequoia Capital — Greg Brockman}

Pragmatic Engineer ~00:00 Mario & Armin: "a good engineer says no, a lot." When AI can produce a confident-sounding rationale for any approach, the senior engineer's "no" gets questioned more — and tribal knowledge gets overridden by AI output that sounds right.

Lenny's ~00:00 Agency isn't evenly distributed. The "skills gap" excuse is gone — what separates builders now is internal conviction and the compounding effect of just starting.

Both clips converge on the same point as topic #13 (NLW) and topic #22 (Brockman): the constraint is now judgment / agency / taste, not capability.

Industry

• YC's $5B+ OpenAI stake under scrutiny — Simon Willison quotes John Gruber: Y Combinator owns ~0.6% of OpenAI, worth $5B+ at the $852B valuation. Simon flags it as hard to verify publicly but worth scrutinizing given YC's dual role as investor and commentator.^{[12]Simon Willison — YC's OpenAI stake}
• The X.400 vs SMTP protocol war — Better Stack short on how SMTP won via simplicity despite X.400's technical superiority; the historical version of every "good enough beats perfect" debate happening in AI infra today.^{[37]Better Stack — X.400 vs SMTP}
• Ferrari's market cap beats Ford + VW + Mercedes combined — Acquired short on luxury scarcity outpunching mass auto by ~160× fewer cars sold.^{[39]Acquired — Ferrari}

AI Models / Tools

• Why distilled models collapse — Nate B Jones short with the geometric explanation for narrow distilled-model capability vs the parent.^{[26]Nate B Jones — Distilled Models}
• OpenAI: Codex preps for sales meetings — Demo orchestrating Google Calendar, Salesforce, Google Drive, Slack, and Gmail from a single Codex thread.^{[32]OpenAI — Codex sales prep}
• Real Python: LLMs as a reverse dictionary — Type a definition, get the word. Genuine LLM-shaped problem.^{[34]Real Python — Reverse Dictionary}

Developer Tools

• React Compiler rewritten in Rust — Better Stack flags Meta's mostly-AI-generated rewrite delivering ~10× faster transformation; still experimental.^{[36]Better Stack — React Compiler}
• deepclaude reroutes Claude Code's tool loop to DeepSeek V4 Pro — Drop-in rerouter, claimed 17× cost reduction. Worth eyeballing if you're paying real money to run Claude Code agentically.^{[31]Github Awesome — deepclaude}
• marimo: docs export to markdown for LLMs — One-click markdown export, send-to-ChatGPT/Claude/Cursor.^{[38]marimo — markdown export}

Hot Takes / Productivity

• Spec-driven dev as a security practice — DeepLearning.AI's comedic skit about shipping AI-generated code without specs and leaking credit cards. Punchline argues for spec-first agentic dev.^{[33]DeepLearning.AI — Spec-Driven Dev}
• Every: how Monologue's landing page was built in Framer — Skeuomorphic dark-mode aesthetic, Paper Shaders for backgrounds (replaces video loops at scale), Rive for stateful interactive animations, footer-as-hero CTA.^{[40]Every — Monologue Framer}