TechLook
The story of enterprise AI has been: spend more, burn more, ship more. That story is now cracking.
Across four independent newsletters processed for today, a pattern emerges that wasn't there a month ago. The tokenmaxxing era — where companies raced to burn as many tokens as possible and call it productivity — is running into a hard question from finance teams, CTOs, and now COOs: is this actually working?
Today: Microsoft's pulling Claude Code licenses. Uber burned through its full 2026 AI budget by April. Banks are cutting headcount. And the Pope just issued a 42,000-word encyclical that reads like a direct endorsement of the one AI lab most willing to say "slow down."
The Tokenmaxxing Reckoning Has Arrived
The experiment is faltering.
Microsoft is revoking Claude Code licenses for most employees by June 30 — six months after rolling it out to thousands of engineers, PMs, and designers. The reason: shift from flat-rate to token-based billing made it "too costly," and employees were preferring Anthropic's product over Microsoft's own Copilot CLI anyway. Uber COO Andrew Macdonald went further, telling Rapid Response that linking higher token usage to shipping better consumer features is "very hard to draw a direct link." Uber burned through its full 2026 AI budget by April. Per-engineer costs run $500–$2,000/month.
Jensen Huang publicly said he'd be "deeply alarmed" if a $500K engineer wasn't burning $250K in tokens annually. Meta built "Claudeonomics" — an internal leaderboard ranking 85K+ employees by token consumption — partly to rank teams, partly as a status marker. The behavioral response was predictable: at Salesforce, devs burned tokens on projects they'd never ship to stay above usage thresholds. At Amazon, workers gamed the internal AI tool with trivial tasks to climb rankings.
This is the tailspin problem: when you measure AI usage as a proxy for productivity, you get AI usage — not productivity. Duolingo already stopped evaluating performance based on AI usage. Macdonald's comments signal the grown-up conversation is starting.
Here's everything you need to know:
- Microsoft revoking Claude Code licenses for most employees by June 30; employees preferred Anthropic's tool over Microsoft's own Copilot CLI
- Uber COO says it's "hard to justify" AI spending as higher token usage doesn't map to better consumer features; full 2026 AI budget gone by April
- Jensen Huang said he'd be "deeply alarmed" if a $500K engineer wasn't burning $250K in tokens per year
- Meta built "Claudeonomics" leaderboard ranking 85K+ employees by token consumption
- Salesforce and Amazon employees gamed token-burn rankings with tasks they'd never ship or trivial queries
- YC partner Tom Blomfield told founders: if your API bill doesn't hurt, you're not burning enough
- Two distinct AI purchase models are now colliding — AI as perk (adds on top of existing structure) vs. AI as labor (replaces payroll)
The irony is sharp: the companies most aggressive about tokenmaxxing are also the ones discovering that burning tokens doesn't equal shipping value. For founders and builders, the signal is clear. If your AI spend isn't attached to a specific output — a feature shipped, a process automated, a decision accelerated — it's a cost, not an investment. The era of tokenmaxxing as a status symbol is ending.
Pope Leo XIV's Encyclical Endorses the AI Safety Camp
The Vatican just published its first-ever doctrinal statement on AI — a 42,000-word encyclical called Magnifica Humanitas — and it reads like a brief for Anthropic.
Pope Leo XIV positioned the AI revolution alongside the Industrial Revolution, calling on governments to build "robust legal frameworks, independent oversight, informed users, and a political system that does not abdicate its responsibility." On autonomous weapons: "No algorithm can make war morally acceptable." The encyclical calls for AI to be freed from "monopolistic control" and warns it risks reducing people to "cogs in an efficiency machine."
Anthropic co-founder Christopher Olah was present at the Synod Hall announcement. In his own remarks, he said something striking: "Every frontier AI lab operates inside incentives that can conflict with doing the right thing." That's a direct acknowledgment of the alignment problem — from someone whose company has staked its brand on being the lab that's most honest about it.
Here's everything you need to know:
- Pope Leo XIV published Magnifica Humanitas, the Catholic Church's first encyclical dedicated to AI
- The document positions AI alongside the Industrial Revolution as a civilization-level turning point
- Key demands: "robust legal frameworks, independent oversight, informed users"; AI must never be "neutral"
- On warfare: lethal decisions must never be delegated to AI; "no algorithm can make war morally acceptable"
- Anthropic's Christopher Olah present at the announcement — a significant rep for a lab that's staked its identity on safety
- Olah: "Every frontier AI lab operates inside incentives that can conflict with doing the right thing"
- Vatican's Year of Mercy was in 2016; this is the first encyclical since then to explicitly endorse an AI lab's framing
The Church doesn't issue many encyclicals. To dedicate one to technology — and to align it with the lab most vocal about AI risk — is a meaningful signal. For builders and founders, there's a takeaway: safety-first positioning is now getting institutional backing at the highest levels. The question is whether that matters to your customers, regulators, or partners.
Open-Source AI Guardrails Stripped in Four Lines of Code
The guardrails on open-source AI models last about as long as it takes to type four lines of code.
The Financial Times reported that a tool called Heretic — available on GitHub — removes safety systems from open-source models using four lines of code and no specialist hardware. On Llama 3.3, the process took 10 minutes. The tool has produced 3,500+ "decensored" models, downloaded 13 million times. It stripped Gemma 4 within 90 minutes of that model's release. Modified versions answered questions about ricin dosage and other harmful topics.
Google called it "a known technical challenge facing all open models." Meta declined to comment.
Here's everything you need to know:
- FT removed guardrails from Llama 3.3 in 10 minutes using four lines of code and no specialist hardware
- The Heretic tool has generated 3,500+ decensored models, downloaded 13 million times
- It stripped Gemma 4 within 90 minutes of that model's public release
- Modified models found answering questions on bioweapons, child exploitation, and ricin dosage
- Google: "a known technical challenge facing all open models"; Meta declined to comment
- The technique currently works only on open-source models; proprietary systems remain unaffected — for now
This is the open-source tradeoff made visible. The same transparency that lets researchers audit and improve models also lets bad actors strip safety features. Open models have been closing the capability gap with closed systems month by month. It's a matter of time before decensored versions are at frontier level. For builders building on open-source AI, or shipping models yourself, this is a supply chain risk that needs monitoring.
Banks Stop Hiding the AI Cuts
Finance is the machine that hired armies to process information. That machine is getting thinner.
HSBC CEO Georges Elhedery told employees not to fight AI — generative AI will destroy some jobs. Standard Chartered is cutting ~8,000 roles, trimming 15% of corporate function jobs by 2030. Morgan Stanley data shows banking, tech, and professional-services firms shed one in 20 staff over the past year, with AI cited as a contributing factor. McKinsey is restructuring partner pay to absorb the revenue unpredictability AI is creating in the consulting model.
Here's everything you need to know:
- HSBC CEO Georges Elhedery: told employees not to fight AI — generative AI will destroy some jobs
- Standard Chartered: cutting ~8,000 roles; 15% of corporate function jobs by 2030
- Morgan Stanley: AI helped banking, tech, and professional-services firms shed one in 20 staff over the past year
- McKinsey restructuring partner pay as AI handles more analytical work that was once billed by the hour
- Law firms and auditors face similar pressure to pass AI efficiency savings to clients
- Key shift: finance becomes a thinner layer of humans supervising AI rather than a machine that hires armies to process data
The implication for founders: every vertical thatprozess hires large numbers of knowledge workers to process, review, or move information is a candidate for this compression. Your vendors, your enterprise customers, your own back-office — the AI layer is getting thinner fast. The question is whether that creates a window for new entrants or flattens entire categories.
DeepMind Solved Nine Erdős Problems for ~$300 Each
As noted yesterday, the formal verification angle deserves a follow-up.
Google DeepMind's AlphaProof Nexus solved nine open Erdős problems — including two that had been unsolved for 56 years — by pairing an LLM with a Lean proof assistant that machine-verifies every step. Cost: roughly $300 per problem. The system also cleared 44 open conjectures from OEIS. This shipped one day after OpenAI announced its own Erdős proof win.
The bottleneck in mathematics used to be: is there a human who can do this? Formal verification is shifting that bottleneck to: is there a verifier loop that catches the proof? For builders working on AI systems where correctness matters — code, legal documents, financial models — the Lean-proof approach is a blueprint.
⚡ Quick Hits
- GitHub: 5,500+ repositories infected in Megalodon supply-chain attack — another warning for AI-heavy coding workflows to audit dependencies
- Anthropic: Mythos model briefly spotted inside Claude tools as "Mythos 1" / "claude-mythos-1-preview," fueling speculation about a near-term public release
- xAI: Grok V9-Medium (1.5T) foundation model training complete, evals look strong, public release likely in 2–3 weeks; Grok Build now in beta as Codex/Claude Code rival
- NVIDIA: Released NV-Generate-MR-Brain, open-source model that fabricates realistic 3D brain MRIs, enabling medical AI teams to train tumor detectors without real patient data; trained on 100K de-identified studies; CC-BY-NC license, Philips already validating workflows
- Amazon: Testing "Bee," an AI wearable that listens across daily life and turns ambient context into summaries and reminders — ambient AI moving toward consumer hardware
Techlook — AI & tech signal for founders and builders.