In Today’s Issue:

🏛️ Dario Amodei’s blueprint for governing the AI exponential

📈 OpenAI preps model 5.6 and an IPO within a year

💸 OpenAI weighs drastic price cuts for the user war with Anthropic

⚡ Google’s DiffusionGemma trades polish for 4x faster text

🧬 CRISPR flips into destroyer mode against cancer

✨ And more AI goodness…

⚡ The Signal

One day after releasing its strongest public model, Anthropic’s CEO asked governments for the power to block models like his own. Dario Amodei’s new essay retires Anthropic’s transparency-first stance and calls for mandatory third-party testing of frontier models, modeled on the FAA. OpenAI spent the same week maneuvering on capital instead: an IPO filing with a listing expected “within the next year,” a new model codenamed 5.6 due this month, and reported plans for drastic token price cuts aimed squarely at Anthropic. Gartner, meanwhile, says data center electricity use will grow 26% in 2026, almost entirely because of AI. The race is widening from models to rules, prices, and power.

All the best,

Kim Isenberg

Google released DiffusionGemma, an experimental open model that generates text through diffusion instead of one token at a time. The 26B Mixture-of-Experts model keeps only 3.8B parameters active, writes 256 tokens per pass, and reaches 1,000+ tokens per second on a single H100, up to 4x faster than comparable token-by-token models; quantized, it fits in 18GB of VRAM on a consumer GPU. Google is upfront that output quality sits below standard Gemma 4, so this one is for speed-critical local work like in-line editing, code infilling, and rapid iteration.

👉 tl;dr: Diffusion is moving from images into open-weight text models, trading some polish for a lot of speed.

Stress-test Amodei’s policy plan against your own job.

🎬 Watch This

Inside Anthropic, the $965 Billion AI Juggernaut (Bloomberg, The Circuit)

– Anthropic reversed a controversial policy that would have secretly degraded Claude Fable 5 for users doing frontier AI research after backlash from researchers who saw it as covert sabotage of competing AI development.

Subscription plans are massively subsidized.

And by massively, I mean absurdly:

SemiAnalysis recently tested this by buying every Anthropic and OpenAI subscription plan and running long-horizon coding tasks until the weekly limits were exhausted, finding that the plans are far more generous than most people assumed. If the price war the WSJ now describes arrives, subsidies of this size are the first thing that gets cut.

Dario Amodei published “Policy on the AI Exponential” on June 10, one day after Anthropic launched Claude Fable 5. For three years, Anthropic’s public policy position was transparency: publish safety frameworks, disclose evaluations, let scrutiny do the work. The essay retires that position. “Frontier AI models, like airplanes, should be required to go through technical testing and auditing, and their release should be blocked or reversed as a threat to public safety if they do not meet high standards of safety,” Amodei writes. His case for urgency is the speed mismatch: AI capability compounds in months while policy moves in years, and in only four years models went from barely writing a coherent line of code to writing most of the code at major AI companies.

The scope goes well past a single regulatory ask; the essay lays out positions across five domains. Regulation: third-party testing above a compute threshold, covering cybersecurity, biological weapons, loss of control, and automated AI research. Economics: real-time measurement of AI’s labor impact, then wage insurance and retention tax credits, then universal capital accounts if displacement proves permanent. Science: pre-approved standards so the FDA and EMA can accept AI-based toxicology and synthetic control arms instead of forcing AI-discovered drugs through 7-to-8-year approval timelines. Civil liberties: a ban on autonomous weapons in domestic law enforcement and closing data-broker loopholes. Geopolitics: a coalition of democracies that shares chips internally, denies them to adversaries, and treats frontier AI less like the internet and more like a strategic capability.

The reception splits along a predictable line. Supporters read it as the first detailed governance plan from someone who actually builds frontier models, backed by Anthropic money behind concrete legislation. Critics note that a certification regime built on compute thresholds, authorized evaluators, and security standards is far easier for incumbent labs to absorb than for new entrants, which lets the safety plan double as a moat. Amodei anticipates the charge, writing that people worry about AI “because they correctly perceive that its risks are real, not because AI CEOs have been insufficiently Panglossian.”

Why it matters: When a frontier lab CEO asks for the power to be blocked, the debate moves from whether to regulate AI to who designs the regime, and that design will decide who can afford to compete.

Sources:

🔗 https://darioamodei.com/post/policy-on-the-ai-exponential

🔗 https://www.yahoo.com/news/politics/articles/dario-amodei-issues-strong-warning-200641485.html

🔗 https://x.com/DarioAmodei/status/2064781775247950326

2026 will redefine IT as a strategic driver of global growth. Automation, AI-driven support, unified platforms, and zero-trust security are becoming standard, especially for distributed teams. This toolkit helps IT and HR leaders assess readiness, define goals, and build a scalable, audit-ready IT strategy for the year ahead. Learn what’s changing and how to prepare.

Download the Toolkit

Most gene editing tries to repair DNA; this approach deliberately destroys it. Researchers at the Innovative Genomics Institute (IGI), spanning UC Berkeley, UC San Francisco, and the Gladstone Institutes, engineered the CRISPR enzyme Cas12a2 to recognize the RNA transcripts of mutant p53, the tumor suppressor gene that fails in roughly half of all cancers and in up to 70 to 90% of some hard-to-treat ones, including ovarian, pancreatic, and non-small cell lung cancer.

The mechanism is a conditional kill switch. As long as a cell carries only healthy p53, nothing happens. When the enzyme senses the mutant transcript, it activates a second function: indiscriminate cutting that shreds the cell’s chromatin and triggers cell death, while healthy neighbors never flip the switch. In a way this takes CRISPR back to its roots: in bacteria, CRISPR systems evolved as destroyers of invading genetic material, not as repair tools.

The work, published in Nature as “Targeting Cancer-Specific Mutations with RNA-Triggered Chromatin Shredding,” showed therapeutic effect in mouse models of lung and liver tumors. The caveats are the usual ones for early-stage gene medicine: delivery into solid tumors is hard, mouse results often shrink in humans, and selectivity has to hold across messy real-world mutations. Still, mutation-triggered cell destruction opens a lane small-molecule drugs never could: targets that were “undruggable” because there was nothing to bind, only something to read.

That investor was wrong. Gamma is now worth $2B, with 50M users and more than half their growth driven by word of mouth.

They're one of 6 AI-native startups in HubSpot for Startups' free Bold Bets Playbook. Replit grew revenue 50x after half the team pushed back on the strategy. Ramp generated 100M+ views from a single stunt. Clay's co-founder wouldn't hang up a sales call until the prospect DMed him in Slack.

Each one took a GTM risk most founders would never greenlight. Each one paid off.

Get the free Bold Bets playbook