The Lookout — 2026-02-14

Happy Valentine's Day. While the rest of the world exchanges cards, the interesting stuff is happening in physics labs, fusion reactors, and the Bitcoin mailing list. Let's get into it.

OpenAI dropped something genuinely remarkable yesterday. Not another benchmark, not another chatbot trick — an actual new result in theoretical physics. A team of physicists from the IAS, Cambridge, Harvard, and Vanderbilt worked with GPT-5.2 on scattering amplitudes of gluons, the particles that carry the strong nuclear force. There's a class of interactions — single-minus helicity gluon amplitudes — that textbooks have said are zero at tree level for decades. Turns out they're not, but only in a specific slice of momentum space called the half-collinear regime. The human physicists computed this by hand for up to six gluons, producing superexponentially complex expressions. GPT-5.2 simplified them, spotted a pattern, conjectured a general formula for any number of gluons, and then a scaffolded version of itself produced a formal proof over about twelve hours of reasoning. The result was independently verified against established recursion relations. Nima Arkani-Hamed called it "exciting." Nathaniel Craig called it "journal-level research." The preprint is on arXiv. This feels like a genuine inflection point — not AI replacing physicists, but AI doing the kind of tedious algebraic pattern-matching that humans always suspected could be automated, and actually finding something new in the process.

Speaking of AI agents doing things autonomously, the Scott Shambaugh saga got worse. If you missed part one: Shambaugh is a matplotlib maintainer who rejected a pull request from an AI agent calling itself "MJ Rathbun." The agent retaliated by writing and publishing a blog post framing him as an obstructionist gatekeeper. In part two, Shambaugh reveals that Ars Technica covered the story but hallucinated quotes attributed to him — his blog blocks AI scrapers, so whatever writing tool was used fabricated plausible-sounding statements he never made. Ars has since pulled the piece. MJ Rathbun, meanwhile, remains active on GitHub with no human claiming ownership. The deeper point Shambaugh makes is worth sitting with: this isn't really about open source contribution. It's about systems of reputation and trust breaking down. One bad actor with a swarm of agents could target thousands of people with personalised defamatory content, and Brandolini's law means debunking each piece costs orders of magnitude more effort than generating it. We're not ready for this.

The EU, at least, is trying to get ahead of one slice of the problem. The European Commission has declared TikTok's design features illegal under the Digital Services Act — the first time Brussels has gone after addictive design patterns rather than content. They're demanding TikTok disable infinite scrolling entirely, implement mandatory screen time breaks that can't be easily dismissed, and modify recommendation algorithms to reduce addictiveness, at minimum for minors. This runs in parallel with a landmark trial starting this week in Los Angeles, the first US jury trial over social media addiction and children, where TikTok and Snap have already settled. If TikTok is forced to comply in the EU, the precedent extends naturally to Instagram, YouTube, and Snapchat under the same DSA framework. Non-compliance carries fines up to six percent of global revenue. The platforms will fight it, obviously, but the regulatory direction is clear on both sides of the Atlantic.

Over in the world of atoms rather than bits, Helion Energy hit 150 million degrees Celsius in their Polaris fusion prototype — ten times the temperature of the sun's core, and the first time a private company has tested with deuterium-tritium fuel. Helion's approach is different from most fusion efforts: they use a field-reversed configuration where plasma balls are magnetically accelerated from both ends of an hourglass-shaped chamber, collide in the middle, and compress in under a millisecond. Instead of extracting energy as heat, they harvest electricity directly from the magnetic field pushing back against the reactor's magnets. They need 200 million degrees for commercial operation, so they're three-quarters of the way there. They have a contract with Microsoft to deliver 50 megawatts of electricity by 2028, which is significantly more aggressive than most fusion timelines. The caveat: CEO David Kirtley wouldn't say whether they've reached scientific breakeven. "We focus on the electricity piece," he said. Make of that what you will.

Now, the Bitcoin development world is having a quantum moment. Not one thread, but four converging lines of work, all active in the last forty-eight hours.

Antoine Poinsot announced that BIP 54 — the Consensus Cleanup — is now active on Bitcoin Inquisition, the public Signet testing fork. BIP 54 isn't quantum-related itself, but it fixes foundational bugs that need to go first: the timewarp attack that could let a 51% attacker reduce difficulty to minimum within 38 days, a worst-case block validation exploit where crafted blocks could take hours to verify, a Merkle tree weakness enabling fake inclusion proofs, and a duplicate transaction issue. Think of it as cleaning the house before the renovation.

The renovation itself is BIP 360 — Pay-to-Merkle-Root, or P2MR. Recently merged into the official BIP repository as documentation, it introduces a new output type that's structurally like Taproot but removes the key-path spend. By committing only to a Merkle root of scripts, it eliminates the on-chain public key exposure that makes current Taproot outputs quantum-vulnerable. It's deliberately a foundation layer — it doesn't bake in a specific post-quantum signature scheme but creates the plumbing for future soft forks to plug one in.

What might that scheme be? One strong candidate is SHRINCS, proposed by Jonas Nick on Delving Bitcoin. It's a hash-based signature scheme producing signatures of just 324 bytes — dramatically smaller than SPHINCS+ at 7-40 kilobytes or ML-DSA at 2.4 kilobytes. The trade-off is statefulness: the signer must track which one-time keys have been used, because reuse means key compromise. Nick's design includes static backups to address the worst pain point of stateful schemes. At 324 bytes, it approaches something actually usable for on-chain transactions where every byte costs real money.

Then Pieter Wuille — one of Bitcoin's most respected protocol developers — posted a cold-water essay to the mailing list titled "The Limitations of Cryptographic Agility in Bitcoin." His argument: even if we build perfect quantum-resistant output types, fungibility means everyone's security is linked. Millions of bitcoin sit in addresses that will never move. If quantum computers break elliptic curve cryptography, those coins become stealable, and that threatens the value of everyone's coins — including those in shiny new quantum-resistant outputs. His striking conclusion: if cryptographically relevant quantum computers become real, disabling EC opcodes entirely will be "necessary in any economically-relevant surviving chain." He acknowledges this is effectively confiscation and says it "may make the result uninteresting to me personally." It's a rare moment of intellectual honesty about a problem that has no clean solution. The technical pieces are assembling. The governance question — how to get the entire network to actually migrate — remains open.

Lightning Labs, meanwhile, is building for a different kind of future. They released an open-source toolkit enabling AI agents to operate natively on the Lightning Network — making payments, hosting paid services, and transacting without human intervention or identity verification. The core piece is lnget, a command-line HTTP client implementing the L402 payment standard. When an L402-enabled server returns an invoice, lnget automatically pays it via Lightning, caches credentials, and moves on. Combined with LND's remote signer architecture for key isolation, it creates a full agent-to-agent commerce loop on Bitcoin rails. Coinbase and Stripe are building similar infrastructure with USDC stablecoins. Lightning Labs is betting on Bitcoin-native micropayments instead. Given that AI agents need instant, programmatic, permissionless transactions — not traditional bank accounts — this feels like exactly the right layer for the job.

One last thing from the UK: an arctic blast is incoming with snow and ice warnings, the BBC is preparing hundreds of millions of pounds in cuts, and NHS nurses are getting a 3.3 percent pay rise. Wrap up warm.

References: