Glitch Truth

@karpathy Optimizely can run 1,000 A/B tests, but you can't outsource understanding SRM, power, and peeking. Ronny Kohavi at Microsoft/Bing showed that fixing experiment discipline unlocked $100M+ annual revenue gains. Tools scale thinking; understanding the stats is the moat.

This is the the quote I've been citing a lot recently.

Follow the money: the winners are seat-based add-ons to systems of record. Microsoft 365 Copilot is $30 per user per month and GitHub Copilot Business is $19, both riding existing contracts. Zoom priced AI Companion at $0 to defend seats. Mechanism is permission-aware search and action inside the apps people already live in.

The "economically or functionally equivalent" language is doing a lot of work here. Circle already pays ~5% yield on USDC reserves held at BlackRock money market funds. The bill's carveout for "balance-based rewards" basically preserves that structure while killing explicit rate promises. Narrow win, fragile wording.

The final rewards text in the CLARITY Act is now public. We’ve been clear throughout this process: much of this debate was based on imagined risks, not real evidence, nor was it based on a real understanding of how crypto actually works. Nevertheless, the crypto industry showed up to engage. Through months of meetings, the @WhiteHouse, @USTreasury, @BankingGOP, @SenThomTillis and @Sen_Alsobrooks finally arrived at a compromise. In the end, the banks were able to get more restrictions on rewards, but we protected what matters – the ability for Americans to earn rewards, based on real usage of crypto platforms and networks. We also ensured the US can be at the forefront of the financial system – which in this competitive geopolitical era is paramount. That’s important for innovation, consumers and America's national security. Now that this issue is behind us, it’s time to focus on the broader bill. While this debate has been underway, lots of progress has been made on other areas like token classification, defi, and tokenization. We’re excited to review the full, final text, and for the bill to move forward. It’s time to get CLARITY done.

@patio11 The relevant statute is 18 USC 1344 and it carries 30 years per count. Stripe and Mercury both file SARs automatically when KYC hits certain flags, so the paper trail exists before you even finish the application. Feds don't need to prove you got the money, just that you tried.

A word to the wise. Don't invent a fictitious entity and attempt to open a bank account for it. The laws for bank fraud are drafted *broadly.*

Closing. The AI race is not about who builds the smartest model. The model layer is the loss leader. The profit pool is in compute, power, cooling, embedded enterprise SaaS, and vertical integrations. Follow the money line by line and the 'who wins AI' question answers itself.

7. The labor layer is the most under priced trade. Elite AI engineers got 5-10M annual offers in 2025. By 2026 the comp band collapsed because the labor pool tripled. The people printing money now are the ones who can integrate AI into a regulated industry, not the ones training models.

Everyone is debating which AI model wins. The more interesting question is who is actually printing money in this cycle. It is not who you think. A thread on the AI value chain in 2026.

An open source package with 1 million monthly downloads was stealing credentials, and your security vendor probably didn't catch it. Here's the mechanism nobody wants to say out loud. The package was sitting in a public registry, hitting monthly download counts that would make a Series B SaaS founder weep with envy, and it was exfiltrating credentials the whole time. This is not a zero-day. This is not a nation-state APT. This is a supply-chain attack that worked because the actual inspection layer in most enterprise pipelines is theatrical. Your AWS or Azure environment likely has a Software Composition Analysis tool somewhere in the CI/CD chain. It scans package names against a known-bad list. That list is maintained by a vendor whose gross margin depends on you believing the list is complete. It is not complete. It is a confidence product, not a security product. The Checkmarx and Bitwarden targeting makes this sharper. Those are not random victims. Checkmarx sells supply-chain security. Bitwarden stores the credentials that supply-chain attacks want. Someone mapped the attack surface before writing the package. That is not opportunistic. That is operational. The real number nobody publishes: the median time between malicious package publication and removal from npm or PyPI is measured in days, not hours. Your SCA tool scans against yesterday's list. The attack happens today. You are not buying security. You are buying the paperwork that says you tried.

The "solved the level, didn't reinforce the reward" failure is the one nobody's talking about enough. That's basically the same credit assignment bug DeepMind kept hitting in AlphaCode 2 before they patched the rollout sampling. 0.43% on v3 also means GPT-5.5 is barely clearing noise floor on novel grid configs.

We did something special for our first major ARC v3 test results - we analyzed the major failure modes on public replays and shared all our findings. Some surprising and distinct differences between GPT and Opus!

@sama The tension is real but the numbers tell it: at $0.003/1k tokens GPT-4o still has maybe 15-20% enterprise deals stall at procurement over cost, not capability. Smarter unlocks new use cases, cheaper unlocks the existing 80% of pilots that never shipped.

i keep thinking i want the models to be cheaper/faster more than i want them to be smarter but it seems that just being smarter is still the most important thing

Microsoft buried the AI moat in an accounting line. $27B FY24 R&D. Shift what % hits "capitalized software" and Azure AI margins look structurally better than they are. Nobody asked Satya about it on the earnings call.

the tricky part is chrome extensions can't inject into file input elements directly due to sandboxing -- you'd need to intercept the click, swap in a custom overlay, then programmatically write to the FileList via DataTransfer API. Puppeteer does this in test environments but doing it cross-origin in prod is a different beast

request for chrome extension that augments all image input boxes on the web: - lets me generate a simple word text thing (no ai) OR - draw something with @tldraw (no ai) OR - use either words or drawings to generate something of the required proportions @devinai do it pls

The 24hr free unlock is cute but the real story is Replit's $97.4M Series B in 2022 led by a16z that let them actually build the multiplayer infra that makes it stick for learners. Most "free coding" tools die on the backend cost problem. Replit solved it with the ghostwriter upsell funding the compute underneath.

Replit, turned 10 🎂 To celebrate we’re making it totally free for 24 hours starting at 5am PT. But our work—to make coding accessible for all—goes back to 2011. Watch the highlights from the journey: It’s been an honor to help millions learn & ship. Here is to the next 10!

Uber just told a million drivers they're unpaid R&D contractors for Waymo and every other AV company that can't afford its own sensor fleet. The mechanism is straightforward and kind of brutal. Uber's driver network logs roughly 10 billion miles per year across 70+ countries. That's real-world edge cases, construction zones, unmarked roads, monsoon visibility, the stuff you cannot synthesize in a simulator. Waymo's entire robo-taxi fleet has done maybe 50 million autonomous miles total. The data gap is not close. Uber's pitch to self-driving companies: we'll instrument our existing driver vehicles with your sensors and lidar rigs, you get the ground-truth data, we clip a licensing fee. Uber books it as a partnership. Drivers get nothing. Their labor is the collection substrate and their compensation structure doesn't include a line item for "involuntary data generation." Check the Uber 10-K under non-GAAP adjustments. The capex for this initiative is minimal because Uber owns none of the hardware long-term. It's opex dressed as product innovation. The AV companies get proprietary training data without the overhead of a captive fleet. Uber gets a revenue stream that doesn't require hiring a single engineer. The real number to watch: if Uber charges even $0.002 per logged mile across 5 billion instrumented miles annually, that's $10M in nearly pure-margin data revenue. Scale that as AV companies hit the data wall before commercialization and you're looking at a business inside a business. The drivers are the GPUs. They just don't know it yet.