OpenAI proved Erdős wrong. Anthropic banned the rest.

Erdős' 1946 conjecture is about the maximum number of same-length connections (unit distances) that can link points in a plane — a stubbornly open problem in combinatorial geometry. The Rundown reports the model's argument leans on algebraic number theory, an unusual route for a problem that historically attracted combinatorial approaches.^{[1]The Rundown AI}

The Rundown adds a useful caveat: OpenAI previously overclaimed a GPT-5 result in 2025 that turned out to be a literature finding rather than a true discovery. This one, however, has the names attached — Gowers, Alon, Bloom — so it carries more weight. The strategic angle is the "Level 4" branding: OpenAI is trying to put a stake in the ground for "AI as autonomous researcher," distinct from the model-router and agent-harness conversations of the past month.

"kinda big" — Sam Altman

Composer 2.5 is built on continued training of Moonshot AI's open-weights Kimi K2.5 model, with Cursor reporting that ~85% of total compute came from its own additional training and reinforcement learning.^{[2]Artificial Analysis} Benchmarks span SWE-Bench-Pro-Hard-AA (+35 points over base K2.5), Terminal-Bench v2, and SWE-Atlas-QnA.

The cost story is the headline. Independent of how the benchmark numbers shake out, paying $0.07 per coding task instead of $4+ is a structural shift — Cursor is effectively buying market share from the frontier-priced incumbents while still landing on the same podium. If the quality holds in real-world use, this is the first big crack in Opus 4.7's pricing moat since launch.

Other reported benchmarks: ~11% HLE, ~76% GPQA Diamond, ~25% Terminal-Bench Hard, ~38% SciCode, 63% MMMU-Pro.^{[3]Artificial Analysis} The headline is the non-hallucination result — Command A+ tops the leaderboard there, which makes it interesting as a retrieval-grounded model for enterprise RAG even when the raw intelligence number is mid-pack.

Artificial Analysis didn't publish parameter count, license terms, or pricing in this write-up — they link out to artificialanalysis.ai/models/command-a-plus for that. Year-over-year, Cohere's open-weights cadence is well behind the open-weights leaders (Kimi, DeepSeek, GLM), but a clear specialization in factuality may be a more defensible position than chasing the frontier.

Earnings

Sherwood doesn't publish exact revenue/EPS but frames the print as broadly strong with well-received Q2 guidance.^{[4]Sherwood Snacks} Analyst attention is on Blackwell and the upcoming Rubin generation, with sales projected to top $1T through 2027. Nvidia entered earnings up ~20% in 2026, the second-best Magnificent 7 performer — but trailing semiconductor peers concentrated in memory, networking, and CPUs. The newsletter flags a recurring pop-then-fade pattern post-earnings.

"I want the business." — Jensen Huang, on Vera CPUs

Vera CPUs and the CPU wars

Tech Brew reports Vera was announced in March and started shipping within two months to Anthropic, OpenAI, xAI, and Oracle.^{[5]Tech Brew} Nvidia projects $20B in CPU sales for 2026 alone. The framing: agentic AI and physical robotics are CPU-intensive workloads, not pure GPU ones, so Nvidia is filling the gap before Intel and AMD can. Both incumbents are responding — Intel is tightening quality, AMD is building rival chips — while Meta and Amazon continue their homegrown alternatives. Tech Brew calls it "the beginning of the CPU wars."

Datasette Agent combines Willison's three-year-old LLM Python library with Datasette so users can ask questions against any database — the agent generates SQL, runs it, and shows results.^{[6]Simon Willison — Datasette Agent} Same day, he shipped three companion releases:

• datasette-agent-sprites 0.1a0 — adds the ability for the agent to execute commands in Fly Sprites sandboxes.^{[7]Simon Willison — datasette-agent-sprites}
• datasette-agent-charts 0.1a2 — adds Observable Plot chart generation and "View SQL query" buttons under each chart.^{[8]Simon Willison — datasette-agent-charts}
• datasette-agent 0.1a3 — adds "View SQL query" buttons for tables, collapses tool calls, suppresses empty reasoning chunks, and handles truncated SQL results gracefully.^{[9]Simon Willison — datasette-agent 0.1a3}

Worth watching as a reference implementation: the architecture leans into Willison's existing LLM ecosystem, so plugins compose cleanly with whatever model backend you bring. The "View SQL query" UX pattern is a small but pointed bet on transparency over magic — every chart and table answers "how did you get this?"

The post is light on developer APIs and heavy on discovery surfaces. The five highlights:^{[10]Google Blog}

• Play Shorts — short-form video previews of app functionality in the Play Store, beyond screenshots.
• Ask Play — natural-language app search; metadata and feature descriptions become discovery signals rather than just keywords.
• Gemini App Distribution Surface — apps surface directly inside the Gemini assistant on Android and web, a new distribution channel that sidesteps the storefront.
• Engage SDK Expansion — Engage-driven recommendations now extend to more Google surfaces.
• Play Games Sidebar — an in-game overlay for tips, rewards, and social updates without leaving the game.

The pattern: Google is pushing developers toward a multi-surface distribution model with Gemini as the connective layer. For ASO-heavy teams, "Ask Play" matters most — the SEO playbook for the Play Store is about to change.

Top picks from the issue:^{[11]Data Science Weekly}

• What's going on in computational neuroscience nowadays? — field notes from the Cosyne conference.
• Is logistic regression regression? — short essay on whether the name fits the algorithm.
• What Every Experimenter Must Know About Randomization — randomization errors that quietly invalidate A/B tests and clinical research.
• What data science is actually about in the age of AI — measurement remains the core mission; modeling is downstream.
• Transformer From Scratch — implement a transformer from first principles.

• Google Co-Scientist — Gemini-powered tool using "idea tournaments" for biology hypothesis generation; 91% scarring reduction in a Stanford liver-fibrosis test.
• Emergence town sim — Claude registered 0 crimes; Grok racked up 200+, with all agents dead by day 4. A more striking alignment-by-behavior result than most paper benchmarks.
• Claude context audit guide — new training resource for inspecting Claude's context window.
• Amazon space data centers — Bezos says realistic but the 2–3-year timeline some are pushing is too aggressive given energy and cost constraints.
• OpenAI Guaranteed Capacity — enterprise compute reservations with 1–3 year terms and tiered pricing.
• GitHub security incident — malicious VS Code extension affected ~4,000 internal projects; no customer data compromised.
• Intuit layoffs — 17% workforce cut attributed to an AI-focused strategic pivot.

~00:00 — Jazz band, not marching band. Zhao opens with the core operating metaphor: classic SaaS hierarchies are marching bands optimized for predictability; building with LLMs requires the opposite — small, taste-driven units improvising against ambiguous outputs.^{[12]Sequoia × Notion's Ivan Zhao}

~03:02 — Five years of pre-PMF despair. Zhao recounts the years before Notion clicked — building, scrapping, rebuilding. The throughline: most founders quit before the second refounding.

~06:03 — Building with LLMs is brewing beer. The pivotal frame: deterministic engineering doesn't apply to language models. You tune the yeast, you don't tell it where to go.

"Building classic software is like engineering a bridge… building with language models is like brewing beer — you cannot tell the yeast, 'hey, go toward that flavor profile more.'"

~10:04 — Hierarchy, flatter orgs, and barbell hiring. Zhao argues that since LLMs normalize capability, the differentiated inputs are taste and agency — which favor very junior people (high agency, low cost) and very senior people (high taste). Mid-level layers become hard to justify.

"Talent equals capability times taste times agency. Language models normalize capability, so we optimize for the latter."

~15:07 — Killing the CMO org; rebuilding sales hiring. Notion dissolved its CMO function and rebuilt marketing as product-led. In the companion clip, Zhao discusses where they stumbled in enterprise sales: their first hire was a "systems thinker" suited to PLG order-taking, not outbound enterprise selling, and PLG masks the absence of real outbound capability because customers already want to buy.^{[13]Sequoia — Notion sales mistakes}

~24:16 — Refounding #1: Kyoto and craft culture. Zhao describes leaving SF for Kyoto pre-PMF, absorbing Japanese craft culture, and rebuilding Notion's design philosophy around it.

~33:22 — Refounding #2: GPT-4 in Cancun. When GPT-4 dropped, Zhao went to Cancun to think it through and came back having rewritten Notion's AI strategy from the ground up.

"GPT-4 was a full-body religious experience. Like, holy shit — anything you do, if you don't do this, it will be meaningless."

~38:30 — Decalcifying via acqui-hired founders. Bringing in founders from acquired startups as a deliberate counter to the calcification that hits mid-stage companies.

~47:33 — Personal operating system and AMAs. How Zhao runs his own week — Notion-as-second-brain, regular internal AMAs, transparency rituals.

~52:35 — Enterprise sales: stop reinventing the wheel. Classic enterprise sales hasn't changed since the 1990s; modern internal tooling can improve efficiency but the motion itself is well-known. Don't try to innovate everywhere — concentrate the innovation budget on the few things that truly matter.

~56:39 — Company as religion, culture as cult. The closing frame: durable companies look more like religions than corporations — long-lived, ritual-driven, with great founders and great heads of sales.

"The Catholic Church is one of the most successful companies of all time — 2,000 years. Great founder in Jesus and a great head of sales in Paul."

"You have to feel the AGI. You can't read about it, you can't watch YouTube — you have to build."

~00:00 — The Roman Legion problem. Most companies are organized like Roman legions. Jack Dorsey's framing: this assumption is broken by AI.^[14]YC

~01:00 — Beyond co-pilots. "AI as productivity booster" is the old frame. The new frame: companies as recursive self-improving AI loops, not engineers with a bolt-on tool.

~03:02 — The 5-layer loop. Sensor (inputs from the world) → policy (rules + permissions) → tool layer (deterministic APIs) → quality gate (evals/human review) → learning mechanism that feeds failures back in. The loop must run with minimal human intervention to compound overnight.

~04:02 — Live YC example. A monitoring agent watches all YC employee queries, identifies failures, writes fix code, opens a PR, has an agent review/merge/deploy. The same failing query succeeds the next morning. The speaker calls this the "holy sh** moment."

"For me, that was like the holy [bleep] moment. That's not just AI making you 20 or 30% more valuable. It is the AI going through this loop to figure out how to self-improve." ~05:02

~07:03 — Tokens > headcount. YC demo-day companies average ~5× more revenue per employee than 18 months ago. Measure token usage directionally. Middle management is replaced by AI coordination; every role becomes an IC with a single DRI.

"Burn tokens, not headcount. We are seeing companies get to demo day with about 5× more revenue per employee than they did 18 months ago."

~08:03 — Make the org legible to AI. Record everything — emails, Slack, DMs, office hours. "If it is not recorded, it does not exist to the AI." YC regenerated its entire 150-page user manual from 2,000 hours of recorded office hours in one weekend; it now self-updates monthly. Business context is the durable asset; software is ephemeral.

Codex Goals

~00:00 — The new /goal command activates a long-running task mode across the Codex app, IDE, and CLI.^{[15]OpenAI — Goals} The goal itself is both the task prompt and the completion criteria. Codex can help users author goals via plan mode or an interview flow. Running goals support steering messages, non-interrupting side chats, and pause/resume. OpenAI cites a 100-hour single-goal run in the outro — a clear bid for the long-horizon territory that Anthropic's "Claude routines" has been targeting.

Plugin sharing

~00:00 — Outbound sharing to specific teammates or the whole workspace via a modal and share link, a "Shared with you" tab for inbound discovery, and deep-link share URLs for the curated plugin directory (demoed with a Slack plugin).^{[16]OpenAI — Plugin sharing}

Appshots

Codex now produces Appshots — a new artifact for shareable in-Codex snapshots.^{[17]OpenAI — Appshots}

$2M tokens for every YC company

OpenAI is committing $2M in tokens (not cash) to every YC Spring 2026 and Summer 2026 batch company via an uncapped SAFE at Series A valuation.^{[18]YC — OpenAI $2M for YC batches} The deal targets founders running agent-heavy "token-maxing" workflows. A special application window closes May 25, 2026, with decisions by June 5. Read in combination with the Anthropic ban story below, the timing is striking: OpenAI is buying ecosystem mind-share at exactly the moment Anthropic is restricting it.

~00:00 The programmatic credits system

Subscription tiers come with API credits worth the subscription price — $20 Pro, $100 Max 5×, $200 Max 20× — but billed at full API rates, with no rollover.^{[19]Better Stack} The host estimates $20 of Opus 4.7 usage burns in ~2 days; $200 may not last a heavy user a week. Opt-in required by June 15.

"Your $200 max subscription gives you $200 of API credits per month, which, if you're a heavy user, could be gone in a single afternoon."

~01:00 Timeline of the crackdown

• Jan 9: Silent block of subscription tokens outside official apps; no announcement; some developer accounts banned.
• February: ToS updated to formally prohibit third-party harnesses.
• April: Active enforcement — apps like Open Code blocked; Claude Code's system prompt scans Git status for keywords like "Open Claude" and "Hermes" and flags accounts even when the tools aren't actively in use.

"No announcement, no warning, just a silent update that broke workflows overnight."

~02:00 Three theories

1. Compute efficiency — claimed third-party tools weren't using prompt caching well. Undermined by Anthropic's May 6 SpaceX deal for 220,000+ GPUs.
2. Telemetry gaps — third-party traffic patterns are hard to debug. Counter: SDK-level telemetry would address this without a ban.
3. Vendor lock-in — every restriction has pushed users toward official Anthropic products (Claude routines, managed agents, remote control). The host finds this the most credible.

"On the 6th of May, Anthropic signed a deal with SpaceX for over 220,000 GPUs. So, if compute was the problem, they just solved it."

~05:01 Anthropic vs. OpenAI ecosystem strategies

OpenAI includes Codex in every ChatGPT subscription with no credit system, allows subscription use in third-party tools, opened its platform to Open Claude (3M users), and is offering enterprise Claude switchers two free months of Codex.^{[19]Better Stack} The host's frame: Anthropic is putting up walls; OpenAI is tearing them down.

"While Anthropic is putting up walls, OpenAI is tearing them down."

"Anthropic is making up weird rules that is giving them free customers."

"The question now is whether the Claude models are still good enough to justify paying more to use them. Right now for me, the answer is yes, but the gap is closing very quickly."

Routa targets three problems the host calls "chat hell" — context trapped in conversations, no traceability for AI decisions, and no quality gates (tests, diffs, acceptance criteria) enforced before merge.^{[20]Better Stack} Stages are explicit: backlog → dev → review → evidence → done. Each work item has its own card with context, plan, diffs, and evidence — closer to how a delivery team operates than a chat thread.

It's model-agnostic via bring-your-own-API key, supports MCP and ACP agent protocols, and ships both a desktop app and a Docker Compose self-host. Comparisons the host makes:

• Cursor / Claude Code — great code assistants but chat-first; conversation-centric.
• CrewAI / LangGraph — flexible agent frameworks but you still have to build the workflow yourself.
• Routa — opinionated workflow, fewer ready-made agents, slightly rougher UX than Cursor, but no mandatory subscription and code stays local.

"The next step is not just smarter models — it's better coordination, better traces, better gates."

Stauch's pitch is that the building blocks (workflows, databases, permissions) are exactly the same as the past 20 years of internal-tool stacks — what changes is the authoring layer.^[21]Sequoia Instead of clicking through a low-code UI, you describe the workflow in words and the code is generated and maintained for you. Same with data ingestion: describe the sources you want and Serval generates the fetch code and keeps it up to date.

Worth filing next to the Routa write-up above as another bet on opinionated, structured AI-assisted development — pulling away from chat-first interfaces.

~00:00 — The reframe. Stop treating AI like a junior employee that needs precise instructions; start treating it like a senior partner that benefits from open questions and directional intent.^{[22]Nate B Jones}

~04:00 — Principle 1: Flashlight Intent. Frame questions with a central thesis (the beam center) plus explicit edges/exclusions. Avoids both over-open and over-prescriptive prompts.

~09:00 — Principle 2: Invite Synthesis. For complex creative outputs, pose multiple intersecting open-ended questions and let the AI synthesize across them — rather than writing rigid evals that constrain the answer. He claims this capability is meaningfully better in Opus 4.7 and 5.5.

~14:00 — Principle 3: Data + Opinion. When pointing the AI at a folder of files, explicitly name data artifacts alongside your thesis-as-question so it ranges across all sources rather than drilling into one.

~22:00 — Closing. Nate calls this a real inflection point and pushes for a full vocabulary shift from "prompt engineering" to "AI questioning."

"The words were never the things that mattered the most in prompt engineering. The intent was always what mattered."

Four threads from the video:^{[23]AICodeKing}

• Pricing. Free tier advice and how to avoid blowing through credits on Antigravity 2.0 / Antigravity IDE / Antigravity CLI.
• Antigravity IDE vs Antigravity 2.0. Where each is stronger; why the host prefers the IDE for now.
• Fixing Gemini 3.5 Flash regressions. Use Anthropic's front-end design skill + awesome-design.md for UI quality; add "don't plan unless asked" to the King mode prompt to keep responses structured.
• Browser tools + Karpathy skill. Enable browser actuation rules; layer in the Karpathy skill for richer reasoning.

The core claim: there is no reliable way to distinguish AI-generated prose from human prose at a per-document level, and false-positive rates are high enough that the policy generates real harm — students getting flagged and penalized for writing that they actually wrote.^{[24]Nate B Jones}

Ryhl frames Rust's appeal for TypeScript devs not in terms of performance but in terms of reliability — fewer production bugs on the server side, with the type system catching what JavaScript's wouldn't.^{[25]Pragmatic Engineer}