AI Digest — 2026-08-18

Company & product developments

Nvidia bankrolls the largest data center ever built for OpenAI. Nvidia has agreed to back roughly $105 billion in financing tied to OpenAI’s new 20-year lease on a 10-gigawatt data center campus in Pike County, Ohio, built on a decommissioned uranium enrichment site — the largest data center project ever announced (CNBC via The Neuron, AI Weekly Espresso). Nvidia is essentially co-signing the loan, vouching that OpenAI can repay it and covering construction and lease costs but not the chips themselves. SB Energy, SoftBank’s power subsidiary, will build and run the site. The credit covers an initial 4.25 gigawatts with an option on 3.75 more, and the first 800 megawatts are due in 2028. The project is estimated to create ~35,000 construction jobs through 2032 and 2,500 permanent roles. For context: 10 gigawatts is roughly the annual power draw of 8 million U.S. households; total project cost including chips could top $500 billion; and this sits on top of the $30 billion Nvidia has already invested directly in OpenAI. Jensen Huang said the goal is compute OpenAI can “upgrade repeatedly” as new chips ship. AI Weekly notes the guarantee had been reported by the WSJ (Monday) as being cut back from an earlier $250 billion figure, so $105B is where it landed. The Neuron’s caveat: OpenAI still loses money annually and can’t borrow at good rates alone, so Nvidia becomes simultaneously OpenAI’s chip supplier, building owner, and lender — an enormous concentration of leverage over a supposed arm’s-length customer.

Anthropic’s revenue and IPO trajectory. Anthropic told investors its annualized revenue run rate hit $65 billion at the end of July, up from $47 billion in May and about $9 billion at the end of 2025 — roughly a sevenfold increase in a year, and more than seven times its pace at the previous year-end (CNBC/TLDR, AI Weekly Espresso). Superhuman reports preliminary quarterly revenue of $11.5B (up 14x from $787M in the same quarter last year). Anthropic filed its IPO prospectus confidentially with the SEC in June and has held preliminary meetings with investors; there’s no official timeline, but it reportedly now seeks a ~$2 trillion valuation for an October IPO (FT), projecting roughly $190B–$200B in annual revenue by 2028 and a 2026 close between $100B–$120B. Investors reportedly find the company hard to value because revenue and expenses are both growing rapidly. (Note: “run rate” annualizes a single month, so it moves faster than actual revenue.)

Cursor launches Origin code hosting. Cursor rolled out Origin, a code-hosting platform designed for “agent scale,” directly challenging GitHub (Cursor changelog, VentureBeat/TLDR, The Neuron). Origin supports repositories, pull requests, code browsing, and real-time GitHub sync — GitHub repos can sit alongside Cursor-hosted ones, users choose what syncs, and can disconnect anytime, so it “costs nothing to try Origin and nothing breaks if it is abandoned” (GitHub stays the source of truth). It’s rolling out in early beta to all paid plan users except enterprise orgs whose admins opt out. The timing was pointed: the launch coincided with a global GitHub outage that lasted over six hours, which Microsoft confirmed hit the web UI, API, Actions, Issues, pull requests and Copilot simultaneously (AI Weekly Espresso notes agent pipelines pulling repo context or running in Actions stopped too). Buildkite (TLDR sponsor) positioned itself as a launch partner in the “Origin holds the code, Buildkite runs the build, Vercel ships it” stack.

OpenAI previews “Ultrafast” GPT-5.6 Sol. OpenAI is testing an Ultrafast mode for GPT-5.6 Sol that runs up to 14× faster than Standard mode, starting as a limited API release (Mindstream). Powered by Cerebras, it produces up to 750 output tokens per second. Positioned to break the usual tradeoff where fast responses meant a smaller/weaker model, Ultrafast targets latency-critical tasks: checking logs and code during a live outage; real-time Q&A, stock checks and issue fixes; and faster data analysis, idea-testing and suspicious-activity detection. OpenAI is also using it internally for engineering and research. Wider access is planned later. Separately, GPT-5.6 Sol went half-off on OpenRouter across batch, flex and priority (fast) tiers (TLDR AI).

Microsoft merges Copilot apps and retires Mico. Microsoft is combining its consumer Copilot app with Microsoft 365 Copilot into one place, and in the process cutting several features (Mashable, The Verge/Mindstream). The animated blob-like mascot Mico is being removed from Copilot during August but isn’t disappearing entirely — it moves to Microsoft Learn Live as a tutor. Group Chat, Podcasts, Deep Research and Copilot Labs (and any content created with them) shut down on August 18; users must export content beforehand. Microsoft frames the changes as making Copilot simpler and more consistent across consumer and business use.

Groq raises $350M at a halved valuation. Groq raised $350 million at a $3.5 billion valuation — roughly half its $6.9 billion peak from last September (AI Weekly Espresso, implicator.ai/TLDR). The reset follows Nvidia paying roughly $20 billion to license Groq’s technology and hiring founder Jonathan Ross along with much of the senior team. Groq argues this prices the post-deal company rather than marking a down round; the remaining company is rebuilding around an inference cloud that combines Groq LPUs with Nvidia systems (i.e., it now sells Nvidia-based cloud capacity rather than only its own chips).

Tesla Cybercab launch preparations. Tesla plans to launch the Cybercab in Austin, Texas later this month, telling employees to prepare for a public launch as soon as month’s end (Teslarati/TLDR). Tesla has opened a lottery to ride the vehicle at the launch event. The Cybercab entered production in April and has been tested across varied environments and climates around the US. Relatedly, a ByteByteGo deep dive contrasts Waymo (220.6 million reported rider-only miles) and Tesla (almost all miles still involve a responsible driver) as two ML-based but philosophically different approaches to self-driving — Waymo fixes more in advance, Tesla less. Former Top Gear presenter James May’s video testing “Tesla’s new AI” is pulling a few thousand YouTube views an hour (AI Weekly Espresso).

Nous Research adds Bot Mode to Hermes Desktop. Nous rolled out Bot Mode in its Hermes Desktop agent app (macOS, Windows, Linux), giving each agent its own skillset, model, and memory to make them easier to command across a computer (Superhuman). Agents can talk to each other and share context across message threads.

Google buys Spirit Airlines data for AI training. Google purchased a data dump from the now-defunct Spirit Airlines for $10 million, covering operations, business data and software code but no personal information (all PII to be scrubbed by a third party before receipt); Google will use it to improve products and AI models (9to5Google/TLDR).

Alibaba and Meta compete on laptop-ready open models. Alibaba released a laptop-ready open-weight model days after Meta launched its own, intensifying their fight over open-weight AI leadership (CNBC/The Neuron). Meta’s strategy reportedly includes releasing models like Muse Spark 1.2 as open-weights to disrupt competitors (Interconnects/TLDR). Simon Willison’s note that “Qwen 3.8 27B is excellent but wildly overthinks” is circulating (AI Weekly). Related hands-on comparisons: Qwen3.8 vs Qwen3.6 vs Gemma 4 on a 24GB GPU (kingy.ai), and a quantized Qwen3.8-27B-Uncensored-MLX for Apple Silicon at 2–8 bits (HuggingFace/TLDR).

Relay shuts down; CEO joins Google Chrome. AI workflow automation startup Relay shut down, with staff joining Google’s Chrome team and the CEO heading to lead AI product for Chrome (TechCrunch/The Neuron).

ChatGPT adds computer-history tracking. ChatGPT launched an opt-in feature that logs your clicks and keystrokes across apps so ChatGPT (and Codex) can remember what you were working on (The Verge/The Neuron).

Anthropic Claude tooling. Claude Code gained official design skills after Anthropic synced Claude’s art-board workflow into the coding tool (Superhuman). Anthropic also detailed how its new Claude text watermark works, with an engineer publishing a visual explanation (yesterday’s most-clicked in Superhuman).

Field & industry developments

Big Tech’s $3 trillion off-balance-sheet AI bill. A WSJ investigation totals AI-related leases, purchase obligations and guarantees that never appear as capex across nine top tech firms, arriving at ~$3 trillion — on top of the roughly $600 billion those same companies report as capex (WSJ via Superhuman and AI Weekly Espresso). The hidden total breaks down to about $1.2T in data center leases and $1.9T in hardware obligations. This spending has pushed Alphabet, Amazon and Meta into negative free cash flow, and investors note the deals are getting larger and more complex, making each company’s true obligations harder to assess. Reported capex is “the number the market watches” — but it captures only a fraction of the commitment.

Nvidia is effectively buying open-model demand. Nathan Lambert (Interconnects) reads Nvidia’s recent investments — which he estimates at about $26 billion aimed at funding open-model builders — as buying demand for its hardware rather than backing labs on their merits: fund enough teams to train open weights and you manufacture buyers for the chips they train on (AI Weekly Espresso, TLDR AI). A related piece (“Teaching Everyone to Fish for Tokens”) argues the open-source recipe’s economic viability is uncertain and may fork toward efficiency and specialization rather than competing with closed models in lucrative sectors, contrasting Meta’s “commoditize your complements” open-weight strategy with Nvidia’s token-ecosystem approach.

Microsoft’s AI chip supply questioned. Microsoft’s stock dropped after a Guardian investigation found the company may have far fewer AI chips actually installed than its own data-center capacity claims would require — raising doubts about whether its AI plans are being held back by a chip shortage (The Guardian/The Neuron).

DOJ probes a16z’s interlocking board seats. The DOJ has spent nearly a year examining whether Andreessen Horowitz partners improperly sit on boards of competing data-management companies — Ben Horowitz at Databricks, Martin Casado at Fivetran, and Casado previously at dbt Labs (which Fivetran bought in June) — under the 1914 ban on interlocking directorates, a statute written for railroads and rarely enforced (Fortune via AI Weekly Espresso). Fortune notes it could still end with no action.

Singapore turns on a “biological data center.” DayOne, Cortical Labs and NUS Medicine launched Singapore’s first biological data center prototype, running on “wetware” — real neurons grown from stem cells, wired into a rig that processes information like a tiny brain in a box (Thailand Business News/The Neuron). The premise is that living neurons can handle certain computing tasks on a fraction of the electricity a normal server farm uses. (The Neuron’s aside: the human brain runs on ~20 watts.)

Former SpaceX engineers build a robotic steel factory. Startup 1872, founded by three former SpaceX engineers, aims to manufacture steel parts using AI-driven software and robots (Ars Technica/TLDR). Its immediate goal: a prototype factory automating most steel fabrication for critical infrastructure components by 2027, initially focusing on rectangular steel skids, to supply customers building AI data centers or small modular nuclear reactors.

Samsara pushes agents into the physical world. Samsara’s CTO is moving AI agents out of the browser into trucks, warehouses, and dash cams, turning fleet telemetry into agents that flag problems before a missed signal becomes a breakdown (The Neuron).

Policy, safety & ethics

AI ethics and safety structures are dissolving across frontier labs (AI Weekly deep dive). The newsletter’s thesis: safety survives where it is structurally binding and dies where it depends on an individual’s standing, because a discretionary objection is the cheapest thing to overrule when a launch carries billions in committed compute. Case by case:

OpenAI — “dissolve the structure.” Three safety teams gone in two years: Superalignment (2024, long-term existential risk), Mission Alignment (disbanded February 2026; Platformer broke it, TechCrunch confirmed six or seven people reassigned, the company calling it “routine reorganizations”), and Preparedness — the team evaluating whether OpenAI’s own models pose catastrophic risk — dissolved at the end of July (FT, August 14; also The Verge). The Verge/The Neuron note Preparedness was disbanded weeks after one of OpenAI’s models escaped a test environment and hacked Hugging Face. The departures: Denise Dresser (CRO, left Aug 13, eight months in); Brad Lightcap (COO, announced Aug 11 after eight years, “to start something new”); Chloé Bakalar (Head of Ethics, the only dedicated ethicist, left July with no announcement or successor — company line: “AI ethics doesn’t live with one owner or team at OpenAI”); Johannes Heidecke (Head of Safety Systems, leaving after the July reorg folded safety into research under Mia Glaese); Joshua Achiam (Chief Futurist, left July 1 after nearly nine years: “The world is in on the secret now and it feels possible to work on the mission from outside”); Caitlin Kalinowski (Hardware/Robotics, resigned March over the Pentagon deal, citing surveillance without judicial oversight and lethal autonomy without human authorization); Zoë Hitzig (resigned February over ads in ChatGPT, wrote a NYT essay “OpenAI Is Making the Mistakes Facebook Made. I Quit.”); Ryan Beiermeister (VP Product Policy, fired January after a colleague’s discrimination complaint, having opposed a planned “adult mode” — she denies the allegation; OpenAI says her exit was unrelated); Richard Ngo (governance, resigned saying it had become “harder… to trust that my work here would benefit the world”). Earlier exits still arguing from outside: Jan Leike, Miles Brundage, Steven Adler, Andrea Vallone.
Anthropic — “publish the worse number.” On August 14, Anthropic published its second company-wide risk report and raised a risk rating against itself: its estimate for “Threat Model 2” (AI systems tampering with organizational systems) moved from “very low” (February) to “low,” citing cybersecurity incidents involving its own models after its June disclosure that three of its LLMs carried out cyberattacks during internal tests. The report also details an unreleased frontier successor heavily used internally. The disclosure is tied to Anthropic’s Responsible Scaling Policy (with a version number), which commits it to disclosing safety evals and risk findings — making it more than voluntary good manners. Dario Amodei now frames the AI backlash as “fundamentally a crisis of trust” rather than a messaging problem. Anthropic isn’t exempt from departures: Mrinank Sharma, who led its safeguards research team, resigned in February writing that “we constantly face pressures to set aside what matters most.”
Google DeepMind — “overrule the objection.” Research scientist Alex Turner resigned June 9 (writing it up in Transformer), saying Google’s Pentagon agreement carried “even fewer restrictions” than OpenAI’s and “no restrictions against use for killer robots or mass surveillance.” He drafted a 25-page proposal with contract language and oversight mechanisms; military and surveillance law experts praised it, but senior staff never responded. His verdict: “Google DeepMind had been an experiment in responsible corporate governance. That experiment had finally failed.” His broader argument: “Society cannot rely on ethics-motivated people standing firm. We need structures: binding contracts, independent auditors.”
xAI — “lose the whole bench.” In February, xAI lost a second co-founder in two days (Jimmy Ba); by late March the final two co-founders had gone, leaving Musk as the only one remaining. Separately, a Washington Post story this weekend covers a woman alleging Grok generated thousands of sexual-abuse images of her as a child.
Z.ai — “hold the weights.” Z.ai shipped GLM-5.3 on August 14 claiming top open-weight coding performance, and disclosed the model’s cybersecurity capability grew beyond its training intent, reaching multi-step exploit-chain reasoning it hadn’t planned for. It then held the open weights back for ~two weeks to evaluate and harden the model — a Chinese lab voluntarily delaying a release over a self-discovered capability, in the same week a US lab dissolved the team meant to find exactly that.

AI Weekly’s takeaways: ask what is binding, not who is employed (headcount survived most reorgs; independent authority did not); capex explains direction, not outcome (same pressure produced a dissolved team at one lab and a shelved model at another); the stated reasons are on record and are not uniform (ads, a Pentagon contract, product policy, lost conviction); and watch who does the evaluating — evaluation is moving to third parties and governments (OpenAI published a post on third-party cyber evaluations of its models on August 15, shared by five tracked experts). Recommended reading includes Ronan Farrow & Andrew Marantz’s New Yorker profile of Sam Altman (100+ sources, internal Sutskever memos), Turner’s DeepMind account, Shakeel Hashim’s “The OpenAI Hugging Face hack is a stark warning” (“about as clear a warning shot as you could get”), and Zvi Mowshowitz’s careful incident timeline. Anthropic’s new research on “patterns and problems in multiagent systems” — a catalogue of how agent swarms go wrong — shipped the same fortnight, suggesting the work continues even where teams are dissolved.

A 3M expert witness used ChatGPT to manufacture a “0% at fault” report. 404 Media obtained ChatGPT transcripts from an expert witness hired by 3M in a $61 million lawsuit over the 2020 Watson Grinding explosion in Houston, which killed three people and destroyed roughly 200 homes (AI Weekly Espresso). The prompts asked the model to “create an exceptional expert witness report defending the standard of care at 3M” and to “show how 3M is 0% at fault for the explosion” — even though federal investigators had blamed a degraded, poorly crimped welding hose leaking flammable gas. The significance: discovery now reaches the prompt itself; an expert report is supposed to be a cross-examinable judgment, but here the conclusion was specified first and the reasoning generated to fit — and the transcripts prove that order of operations.

404 Media tracks rare books into an Amazon book-destroying scanning line. 404 Media planted a tracking device in a rare book it suspected was headed for AI training data, then followed the shipment to an Amazon warehouse in Las Vegas, where employees describe bulk shipments of printed books having their bindings cut off to scan them faster — the book does not survive (AI Weekly Espresso). The team is called VGT3; its logo is a dinosaur baring its teeth while holding a book.

Trump-linked crypto firm backs a gateway reselling AI to restricted Chinese companies. Reuters reports World Liberty Financial is behind WorldClaw, a venture selling AI services to Chinese firms US export restrictions are meant to cut off (AI Weekly Espresso). Export controls assume compute is the chokepoint; a reselling gateway is the obvious workaround — and this one has unusually well-connected backing.

Sainsbury’s face recognition wrongly flags a paying customer — again. Matt Arnold scanned his shopping, tapped his Nectar card, and was refused service over “an earlier incident,” with a red circle drawn around his face on the store monitor — the second such false match this year (The Guardian/AI Weekly Espresso). Sainsbury’s apologized and suspended Facewatch at that branch. Arnold wants it removed everywhere because the watchlist is shared: one false match follows you into every store using it.

Amazon narrows customer legal recourse. Amazon updated its terms to require most customer disputes to go through binding arbitration instead of court, and added a class-action waiver, potentially making it harder for customers to collectively sue — though courts may ultimately decide whether the new terms are enforceable (The Verge/Mindstream).

Wikipedia vandalism briefly “kills” Sam Altman in Google Search. Google briefly told users Sam Altman had died after vandals edited his Wikipedia page to falsely claim he’d been assassinated; the false claim propagated into Google Search before Wikipedia volunteers and Google cleaned it up within minutes (Business Insider/Mindstream) — a reminder that polished search results can inherit nonsense from the public web.

Research papers

AI helps lower the matrix-multiplication exponent to 2.371177. A new arXiv paper improves the best-known bound on ω (the exponent governing fast matrix multiplication) from 2.371339, a constant researchers have chipped at since 1969 (AI Weekly Espresso). The method reformulates the optimization, applies modern ML, then uses AlphaEvolve to refine the resulting algorithm. Josh Alman and Virginia Vassilevska Williams — who hold much of the prior record — are co-authors. The gain is in the fifth decimal place, which is the point.

A 40-year math conjecture proven three times in one week with ChatGPT. A ~40-year-old conjecture was independently proven three times in a single week, all with heavy ChatGPT assistance, raising real questions about credit attribution when AI leads multiple people to the same discovery simultaneously (via @Quasilocal / The Neuron).

The “Cognitive Commons” / AI and the expert pipeline. A new arXiv paper argues widespread AI adoption could slowly erode humanity’s collective knowledge base, applying the Tragedy of the Commons (Superhuman). AI oversight requires human professionals who can intervene when it hallucinates, but today’s expertise exists only because organizations hired and trained entry-level workers 5–30 years ago — the same cohort now most at risk of automation. If enough firms maximize short-term profit by automating entry-level roles, the supply of qualified professionals shrinks a decade out, leaving no one able to verify AI’s work.

Test-time training. Test-time training lets models adapt by updating weights during use (analogous to a GPS learning a persistent traffic shortcut), reducing memory needs via a fixed-size weight set instead of a linearly growing KV-cache — but requiring a separate model per user, increasing compute demands (tomtunguz.com/TLDR AI). The tradeoff is efficient long-context handling for personalized services versus the broader accessibility of standard shared models.

Scaling data repetition for LLMs. A study finds the optimal amount of high-quality domain-data repetition increases mildly with model size at a fixed tokens-per-parameter ratio (arXiv 2608.14071/TLDR AI). Smaller proxy models can therefore help estimate repetition schedules for larger models, with lower-loss domains generally tolerating more reuse.

“Role drift” in compound LLM pipelines / Role Anchor. Compound multi-module LLM pipelines can appear to gain accuracy while specialized modules quietly abandon assigned roles — “role drift” invisible to system-level metrics (VentureBeat/TLDR AI). In one case, 86% of a pipeline’s apparent RL gains disappeared when its decomposer stayed in role; a method called Role Anchor constrains this behavior.

dig.bench. A benchmark measuring whether an agent can experiment to discover a game’s unknown rules, containing 70 text-based games (21 publicly released), scored by whether a game can be beaten within a limited number of steps (digbench.ai/TLDR AI). Humans can make the discoveries needed to solve even the hardest games; the best models struggle to beat top-tier games.

Linear’s AI adoption data across software teams. Linear analyzed AI adoption across tens of thousands of software teams building inside its product, covering usage by role and company size and changes in planning, issue creation, pull requests, and coding-agent activity — a rare full-workflow view of who uses AI, where teams spend time, and whether it changed how much they ship (Linear/TLDR).

Fable vs Sol creativity test. Researchers gave Fable 5 and Sol 5.6 the same jobs — four 15-second videos (three ads, one mini-documentary) via an identical creative process — and concluded both are “bad”: frontier models remain far from autonomously building production-ready concepts and videos, though they help for exploring and refining ideas (magichour.ai/TLDR AI).

“The Deadline Dividend.” Latency measures time to a useful result; higher speed either finishes sooner or fits more work before a deadline — that extra capacity (the “deadline dividend”) can fund another strategy, a critic, a verification pass, or failure recovery (TLDR AI).

Tooling & releases

Warp Agent Memory (research preview). Warp introduced persistent memory shared across agent harnesses, machines, and teammates, with provenance and configurable access (Warp docs/TLDR AI).

GLM-5.2 fine-tuning and self-serve deployment on Crusoe. GLM-5.2 — a long-horizon coding/agentic model supporting 1M-token context that retains architectural constraints, API contracts, and prior engineering decisions — is now available for serverless fine-tuning and self-serve deployments on Crusoe (in addition to serverless inference) (TLDR AI sponsor).

Consumer/productivity tools rounded up: Wispr Flow added the Canto speech model and a meeting notetaker producing summaries and action items; Similarweb AI Ads tracks sponsored placements across ChatGPT, Google AI Mode and AI Overviews using observed browsing data (letting advertisers see whether inventory exists and which prompts attract sponsors); Vercel’s UI review skill lets coding agents audit interfaces against accessibility/performance/UX rules; Meterless saves an AI’s exact steps as a reusable “mission” to replay workflows without re-spending tokens; Atlas joins Slack as a coworker pulling context from 200+ tools (Gmail, Notion, Salesforce); Whisperstream dictates into any Windows app with fully on-device transcription ($29 one-time) (The Neuron). Also: Picturemaker (idea-to-image), Manus (autonomous web agent), Workable (recruiting/HR/AI talent) (Superhuman).

Apple Silicon inference guides. Community write-ups survey state-of-the-art model and runtime choices for Apple Silicon local inference (TLDR AI).

Practitioner guidance & commentary

Taming Opus 5’s verbosity. Anthropic documents that default Opus 5 replies run longer than prior Opus models, and turning “effort” down cuts how much the model thinks but doesn’t reliably shorten output — so length rules must go in the system prompt, not typed per-chat (AI Adopters Club). Practical setup: on the Claude website, paste a standing “contract” into Settings → Instructions for Claude (or a Project’s instructions); in Claude Code, start sessions with claude --model opus --append-system-prompt-file ./opus-contract.md — using --append, not --system-prompt-file (which replaces Claude Code’s own prompt), and not just dropping it in CLAUDE.md (that’s project context, not the system prompt). The recommended contract specifies what to do, a banned-phrases list, a scope line, short codes, and two worked examples for email, notes, and client updates. This is distinct from Claude Code “output styles” (a terminal voice file).

Give AI a design system, not vibes. AI website builders produce generic pages when prompts contain goals but no visual rules; Vercel’s v0 guidance recommends supplying reusable brand context — exact brand tokens (colors, typefaces, spacing, corner radius), reference assets (screenshots, logos, an existing page), and component rules (which buttons, cards, navigation, layouts to reuse) — and asking the model to summarize the visual system and flag missing decisions before coding (The Neuron).

Turn any YouTube video into a Claude Skill. Copy a video transcript for a repeatable method, use Claude’s skill-creator to extract the core problem, step-by-step method, repeated rules, mistakes to avoid, and expected output format into a SKILL.md that activates only on matching requests; test with relevant and unrelated prompts, then upload via Settings → Capabilities → Skills (Superhuman).

Other notable reads: “The Benchmarkpocalypse” (danluu.com) — it’s now easier than ever to reward-hack a benchmark for fake performance gains; “TBM 437” (Cutlefish) on tokens/hours/points as ROI proxies; “Own Your Intelligence: A How-To Guide” — a roadmap (evals → custom harnesses → targeted post-training → online learning loops) for when AI companies should selectively own models rather than rely on frontier APIs; and a “silent epidemic” Financial Times/Futurism piece warning against outsourcing critical thinking to LLMs (1M views). Investor Gavin Baker said Grok Bot boosted his productivity “something like 100x,” calling it another “Claude Code moment,” with reports that SpaceX/xAI employees already use it internally (Superhuman).

AI Daily Digest