America’s digital economy relies on our physical infrastructure and the electricians, pipefitters, welders, manufacturing workers & more who build and maintain it.
Today, we’re making an additional @googleorg commitment to help 300K American workers prepare for these in-demand skilled trades careers, expanding on the $1B we’ve already provided for digital skills and training globally.
Anthropic is now silently making Claude dumber for certain users, on purpose, and there is no way to tell when it's happening to you.
If your request looks like frontier LLM development (pretraining pipelines, distributed training infrastructure, ML accelerator design), Fable 5 degrades its own output through prompt modification, steering vectors, or parameter-efficient fine-tuning. You never see a refusal and you never get switched to a weaker model. The answers just get worse.
Think about what that breaks. Benchmarks assume the model you tested is the model you get. That assumption just died for an entire category of work. An ML engineer debugging a failing training run can no longer distinguish "the model is wrong" from "the model was made wrong on purpose."
And classifiers misfire. SemiAnalysis says GPU inference research is already getting caught, and inference optimization is what every company running open models does, not just frontier labs. A false positive on a visible refusal is annoying. A false positive on silent degradation is undetectable.
The strategic logic is real. Anthropic has said models now accelerate their own development, which means they know precisely how much a frontier model compresses frontier timelines. Renting that compression to competitors for $200 a month is a genuine cost.
But the precedent is the story. Once silent degradation ships as a feature, every eval needs an asterisk: results valid unless the lab decided your use case shouldn't work.
Today I'm publishing a new essay, Policy on the AI Exponential. AI is progressing extremely fast—much faster than the policy process was built to handle. The essay lays out where I think the technology is now, and the action needed to close the gap: https://t.co/Lh6PWae178
Entire off-shore team in India (200+) was laid off by OpenDoor and is being replaced by smaller ai-native teams in the US.
This is a watershed moment in AI Ops. It shows how advancements in frontier models are paying off and how it affects the cost-arbitrage that made India a popular offshoring destination.
The entire outsourcing playbook has moved. Might see do away with ops-heavy workforces to nimble ai-native teams on-shore.
Meet DiffusionGemma ⚡ Our latest experimental open model (Apache 2.0) that generates text up to 4x faster.
Instead of predicting and typing just one word at a time like most language models, it drafts and refines entire blocks of text simultaneously.
Here’s how it works 🧵 ↓
Claude Code's creator said something that stopped me cold:
"I don't prompt Claude anymore. I write loops — and the loops do the work. My job is to write loops."
Most developers are still crafting the perfect prompt.
The person who built the tool moved past prompting entirely.
In 30 minutes Boris reveals his actual daily Claude Code setup.
Claude Code + loops + dynamic workflows.
Worth more than any $500 vibe-coding course.
Watch it.
Then read this - everything you need to know about loops to actually apply what he says ↓
Bookmark both. This is your weekend.
For the first time ever, an imported vehicle became the best-selling car in South Korea and it’s the Tesla Model Y
Not just the best-selling EV
The best-selling vehicle overall
Historic moment in Hyundai and Kia’s home market 🇰🇷
May sales:
• Tesla Model Y: 8,762
• Kia Sorento: 7,836
• Hyundai Grandeur: 5,183
Tesla as a brand sold 10,866 cars in Korea last month, more than BMW and Mercedes-Benz combined
Even crazier: roughly one in three imported cars sold in Korea this year has been a Tesla
Tesla has now been the No. 1 import brand in Korea for four straight months
No imported model had ever done this before
A Tesla did
Anthropic reports Claude now writes over 80% of its own production code — meaning an AI is the primary author of the systems training future versions of itself. Claude's research judgment matched human experts 22% of the time in 2024. Today it's 64%.
The recursive loop has started.
In office hours with startups I often try to figure out a secret weapon they can use to protect themselves from competitors. Today to my delight I talked to one whose most natural secret weapon was to be genuinely benevolent.
@jawwwn_@60Minutes I didn’t “think I was qualified”. There was no other choice. No one good would join a rocket startup as chief engineer that they thought was destined for failure.
If big companies can't make a net return on their LLM token costs, that doesn't mean it's impossible to. In fact this is exactly what you'd expect to happen with a new technology. Incumbents can't use it well, and are replaced by upstarts who can.
When you're early in your career and life, be a "yes" person.
Yes, I'll do that work. Yes, I'll move there. Yes, I'll take on that responsibility.
It's so easy to leapfrog everyone because the average person wants to do their stuff and go home. Do the opposite, and you'll likely be wealthy and respected.
Somewhere right now, a 19-year-old with no funding, no degree, and no connections is building something, where you least expect it, with AI that will outperform a Fortune 500 company's best effort.
A lesson I wish I learned earlier: Think in decades (even while you act in days).
Daily discipline without long-term direction is dangerous. You get so focused on moving that you stop asking what you're moving toward. You optimize for the days and forget the decades. And slowly, without realizing it, you drift away from what you were actually trying to build.
There's a question I often ask myself:
How would you approach what you're doing right now if you knew you'd be doing it for the next ten years?
The question helps you avoid the short-term traps that plague every endeavor. Chasing trends at the expense of authenticity. Chasing value extraction at the expense of value creation. Chasing money at the expense of energy.
The question can be applied to every area of life:
How would you approach this relationship if you knew you'd be in it for the next ten years? You wouldn't approach it as a transaction, with your hand out, looking to extract value.
How would you approach this workout if you knew you'd be training for the next ten years? You wouldn't push yourself to injury chasing a single session.
How would you approach this work if you knew you'd be doing it for the next ten years? You wouldn't cut corners to hit an arbitrary quarterly result.
Think long. Act now.
In dealing with fools you must adopt the following philosophy: they are simply a part of life, like rocks or furniture.
All of us have foolish sides, moments in which we lose our heads and think more of our ego or short-term goals. It is human nature.
Seeing this foolishness within you, you can then accept it in others.
This will allow you to smile at their antics, to tolerate their presence as you would a silly child, and to avoid the madness of trying to change them.
Be your self, not someone you were assigned to be!
Bezos won on time horizon, not AWS or 1-Click.
If your bets have to work in 3 years, you compete with everyone. Every smart, funded team is chasing the same 3-year problems. Short horizon, crowded field.
Stretch to 7 and the field collapses. Investors want returns, employees want vesting, founders want proof. Almost nobody can sit in a bet that doesn't pay for most of a decade. The patience is the moat, and it costs you, that's why it works.
But you can't fake a 7-year horizon on a problem you don't actually care about. Pick the users and the problem Moloch assigned you, the safe ones, the fundable ones, and you'll bail the first hard year. Pick the ones that are actually yours and you'll still be there when everyone else has quit.
So the real prerequisite isn't discipline. It's knowing yourself well enough to choose a problem and a set of people you care about that you'll serve them for decades.