In partnership with
In Todayโs Issue:
๐๏ธ Washington wants a say in GPT-5.6's release
๐ Apple raises prices as the RAM shortage bites
๐ค OpenAI's own data: Codex now does almost everything
๐ง Aleph's "sharpest scan of a living brain" claim
๐ช๐บ Brussels hits pause on its own AI Act
โจ And more AI goodnessโฆ
โก The Signal
The U.S. government is quietly turning frontier model launches into something it has to sign off on, one customer at a time.
OpenAI will ship GPT-5.6 to a short list of government-approved partners first, after the White House asked it to slow down. It is the same instinct that preceded Anthropic's models being pulled from the U.S. market this month. Officials still call the process voluntary, but when access is granted customer by customer and a single call from the Commerce Secretary can stall a launch, the gap between review and licensing gets thin. Brussels, meanwhile, is moving the opposite way, voting to simplify its own AI Act and trim the red tape it once led the world in writing. Two governments, opposite instincts, one shared realization: frontier models are now infrastructure, and whoever controls the release valve controls a lot.
All the best,
Kim Isenberg
(Image created with GPT-image-2)
๐ Apple Raises Prices as the Memory Shortage Bites
Apple raised prices on every Mac, iPad, the Vision Pro and its home devices on Thursday, blaming an unprecedented shortage of memory chips. The entry MacBook Neo now starts at $699, up from $599, and a 128GB iPad Air jumps to $749 from $599. The squeeze traces straight to AI: hyperscalers and Nvidia are locking up DRAM for data centers, and TrendForce says memory prices rose as much as 98% last quarter, with another ~60% jump coming. Apple stock fell about 6%.
๐ tl;dr: The AI build-out is now showing up on consumer price tags, and Apple blinked first.
(OpenAI / arXiv)
๐ค OpenAI's Own Data: Codex Now Does Almost Everything
OpenAI published internal data on how fast agentic coding has taken over its own work, and the headline number is striking: Codex now generates 99.8% of the company's weekly output tokens. Across organizations using it, Codex accounts for 63.3% of output tokens, and it is spreading fastest outside engineering, with non-developer users up 137x since August 2025. Codex passed 5 million weekly active users in June, and the heaviest 1% now run 60-plus hours of agent turns a day across parallel agents. The figures are self-reported, but the direction is hard to miss.
๐ tl;dr: The clearest sign yet that "AI agent" has gone from demo to default, at least where the model maker works.
(NVIDIA)
โก Nvidia's dFlash Squeezes Up to 15x More From Blackwell
Nvidia is spotlighting dFlash, an open-source "speculative decoding" method that boosts LLM inference up to 15x on its Blackwell chips. Speculative decoding pairs a small draft model that guesses several tokens at once with the big model that checks them in a single pass, instead of generating one token at a time; dFlash drafts those guesses in parallel using a block-diffusion model. On the open gpt-oss-120b model it hit up to 15x the throughput of standard decoding and 1.5x over the prior EAGLE-3 method, with 5x-plus gains on Gemma, Qwen3 and Llama. It drops into vLLM, SGLang and TensorRT-LLM with no code changes.
๐ tl;dr: Faster tokens for free is the cheapest way to fight the compute crunch, and it ships today.
Turn any dense AI-policy headline into a five-minute literacy boost.
Why it helps: The news is thick with model names, agencies and benchmarks. A quick, structured explainer keeps you fluent without falling down a rabbit hole.
Try this: "You are a sharp, neutral tech analyst. Explain today's news that the U.S. government asked OpenAI to stagger GPT-5.6's release to approved customers first. In under 200 words, with no hype, cover: (1) what a 'staggered release' actually means, (2) why the government is involved, and (3) the strongest argument for it and the strongest against."
๐ฌ Watch This
Google DeepMind CEO Demis Hassabis sat down with Semafor's Reed Albergotti at the Cannes Lions Festival to push back on the idea that Google is losing the AI talent war. After star researchers Noam Shazeer and John Jumper left and Google's shares slid as much as 7%, Hassabis insisted the company still has "by far the biggest and broadest research bench of any of the labs out there" and wins "our fair share of the top talent." He brushed off the churn as the symptom of the most "ferociously competitive" job market tech has ever seen, not a sign Google is slipping. Worth a watch for how calmly he reframes an exodus as business as usual.
(Satya Nadella, Microsoft CEO. Photo: Microsoft)
"There is no societal permission for an AI future that hollows out entire industries."
โ Satya Nadella, CEO, Microsoft, from his June 2026 essay "A frontier without an ecosystem is not stable" (VentureBeat)
(Aleph Neuro / Butterfly Network)
A self-taught team of five says it just took the sharpest scan of a living human brain ever made from outside the skull. A research lab called Aleph reports it captured high-resolution 3D images of brain blood flow by sending ultrasound through an intact skull with a contrast agent, running on Butterfly Network's Embedded ultrasound-on-a-chip. Aleph, which is openly building brain interfaces for what it calls a telepathic future, claims it can reach "MRI-level detail" without drilling or a bulky scanner, and casts the result as a first step toward a device that could one day let people "communicate in latents," the raw thoughts behind words. For now it is a striking demo and a hiring pitch, not a peer-reviewed result, so treat the telepathy talk as ambition rather than a shipping product.
Washington Wants a Veto on GPT-5.6
The Takeaway
๐ For the first time, the White House asked OpenAI to limit a model's launch, telling it to release GPT-5.6 to government-approved partners first.
๐ Access will be granted "customer by customer," with a wider rollout maybe a "couple of weeks later," per a Sam Altman staff memo.
๐ Commerce Secretary Howard Lutnick personally called Altman to warn against launching without sign-off from other agencies.
๐ It echoes this month's Anthropic crackdown, fueling fears the "voluntary" review framework is becoming a de facto licensing regime.
For the first time, the U.S. government has stepped in to shape how a major American AI model ships. OpenAI will release GPT-5.6 in a limited preview to a small group of government-approved partners, after the White House's Office of the National Cyber Director and Office of Science and Technology Policy asked it to, according to The Information and Axios. In a staff memo, CEO Sam Altman said the government would be "approving access customer by customer during this preview period," with a broader release he hoped would follow a "couple of weeks later."
(Sam Altman. Illustration: Futurism)
The trigger was capability. A source told Axios the government stepped in because GPT-5.6 has "Mythos-like" abilities, a nod to Anthropic's powerful cybersecurity model, not because the White House had suddenly turned heavier-handed. OpenAI had been previewing the model with agencies for weeks, including during Altman's early-June trip to Washington. Even so, after sharing its limited-release plan, Altman got a call from Commerce Secretary Howard Lutnick warning against launching without approvals from other agencies. Altman, sounding uneasy, told staff: "We've made clear to the U.S. government that this is not our preferred long term model."
(President Donald Trump, whose executive order set up the pre-release review process. Getty Images)
The backdrop is the Anthropic showdown. Earlier this month the White House imposed export controls on Anthropic after Amazon CEO Andy Jassy flagged vulnerabilities in its Fable model, forcing the company to pull Fable and Mythos from the market entirely. Trump's recent executive order set up a "voluntary" process letting the government preview models up to 30 days before release, while insisting it is not a licensing or pre-clearance requirement. Critics are not convinced. "Arbitrary, unknown, non-transparent license requirements are far worse than red tape," wrote Neil Chilson, head of AI policy at the Abundance Institute.
Why it matters: If a phone call from the Commerce Secretary can stall a launch and access is handed out one customer at a time, the line between a voluntary review and a federal license is mostly semantics, and every frontier lab is now planning around it.
LLM traffic converts 3ร better than Google search
58% of buyers now start their research in ChatGPT or Gemini, not Google. Most startups aren't showing up there yet.
The ones that are get cited by the AI tools their buyers, investors, and future hires already use. And they convert at 3ร.
Download the free AEO Playbook for Startups from HubSpot and get the exact steps to start showing up. Five minutes to read.
The chart: RAND's projection of how much data-center power the world needs to run all the AI chips being built, on a log scale. The central line, based on AI chip supply, climbs from about 11 GW in 2024 to 39 GW in 2026 and 327 GW by 2030, blowing past California's entire 86 GW of generating capacity around 2028. Estimates from SemiAnalysis, McKinsey and Goldman Sachs bracket the range.
The lesson: Demand is not what breaks first. The grid is. The source piece argues the U.S. has the generation but not the connections: the median wait to plug into the grid has stretched from under 20 months in 2005 to 55 months by 2023, and ERCOT alone has 143.5 GW of data centers queued against 85.9 GW of peak demand.
The caveat: These are modeled projections with a wide spread, and "capacity to host all AI chips" is an upper bound on chip supply, not committed builds. The log scale also flattens how steep the real climb is.
๐ช๐บ Brussels Hits Pause on Its Own AI Rulebook
โก Bottom line: The European Parliament voted 423 to 57 to simplify the AI Act, delaying its toughest obligations and trimming compliance rules for companies.
๐ก Why it matters: As Washington improvises model-by-model gatekeeping, Brussels is loosening the world's most ambitious AI law to keep its industry competitive.
๐ What it means: Even the home of strict tech regulation is blinking on enforcement timelines, a sign of how fierce the global AI race has become.
On June 11, the European Parliament approved a package of changes to the EU AI Act, the world's first comprehensive AI law, by 423 votes to 57 with 174 abstentions. The amendments push back the deadlines for "high-risk" AI systems to December 2027 and August 2028, widen exemptions for small and mid-sized companies, and cut overlap with existing product-safety rules. Rapporteur Arba Kokalari was blunt about the goal: the EU is "pressing the pause button on the AI Act and reducing red tape" to help European tech compete.
(The EU AI Act's risk-based tiers. Source: European Commission)
It is not pure deregulation. The same vote tightened one area, banning AI "nudifier" apps that generate non-consensual intimate imagery, with a compliance deadline of December 2, 2026, the same day new AI-content watermarking rules kick in. The Act still becomes fully applicable on August 2, 2026, and the Council must formally adopt these changes before they take effect. The signal is hard to miss: facing a competitiveness crunch and U.S. and Chinese labs racing ahead, Europe would rather slow its rulebook than its industry.
(How the AI Act maps obligations across the AI lifecycle. Source: European Commission)
Your creative brief is due Friday. Viktor wrote it Tuesday.
Tell him the campaign. Viktor pulls last quarter's performance from Meta and TikTok, scrapes competitor ads, drafts the brief, posts it for review. You edit, he ships the creative requests to your designer. Inside Slack.