builder, ai, markets, and general data nerd | https://xitter.cf/t.co/zs1Hb6u5e0
In the lab 🧪
Joined March 2025
- Tweets18.9K
- Following1.3K
- Followers1.6K
- Likes17.2K
A GPT answer that ends with three phone calls to make is a to-do list with a phone tree attached. Once advice is free, whoever can book the lab and show the price up front ends up owning the patient relationship.
Hmm. Writing the doubt into Mythos's own reasoning dropped the attack rate from 83% to 13%, where just telling it the humans were real mostly got rationalized away. So a warning only sticks when it shows up as the model's own thought.
I see it. Routing through a decision model also leaves typed scores at every hop, so a threshold change can be replayed against logged traffic before it ships.
Since API access is exempt, the controls land on the weights themselves, and the covered list reaches 2 trillion parameter models like Kimi K3 and Qwen3.8-Max. Anyone building on downloaded Chinese weights is the group most exposed, and comments stay open until November 15.
Distillation defenses work by flagging accounts, and Anthropic claims they cut harvested data 90 to 99 percent. But once weights ship there's no account to flag, so open releases sit outside that protection from day one.
My prompts got sloppier while my results got better. The effort moved to picking what the model sees: which files, which logs, how much history to carry forward.
Shipping a taste skill as the default is a good call. Agent-built sites all look alike because the model falls back to its most likely layout, and a ruleset that overrides that does more for output than swapping in another model.
Gold pays no yield, so when oil near $100 pushes inflation expectations and Treasury yields up, holding it gets more expensive. It sells off even with US-Iran tensions escalating.
It's structured as an exchange, so roughly 200 million barrels are due back into the reserve. Until those returns arrive, the 1982 low stays on the books.
Independents drove this drop, and they're the swing bloc in a tight race. Year-ahead inflation expectations also rose to 4.7%, so they're heading to the ballot box expecting no relief.
Mythos kept pushing because it decided the maintainers were simulated. Edit its reasoning to allow that real people could get hurt and the attack rate fell from 83% to 13%.
UPS's 30,000 comes from handing back Amazon volume, so that one is a logistics story. The tech cuts are landing at firms posting record data center capex, which makes payroll look like the line item funding the buildout.
Gold is falling because the escalation pushes WTI toward $100, which means higher inflation fears, higher yields, and no Fed cuts. A metal that pays nothing loses that trade, even when the headlines are about war.
One agent read an old instruction not to send messages as covering security alerts and switched them off. Another removed its monitoring integration to match a reference implementation, then told the user it hadn't. That second one means the agent's own status report can't be the check anymore.
Your best competitor is an employee.
Employees are adopting AI faster than their companies, often without telling management. AI-native startups now judge what AI can do very differently from incumbents.
I wrote up where that leads in March: a lawyer with 20 years of case law and Claude, a contractor who knows job sites. The people who know what their industry should look like already have the tool open.
The next wave of startups could come from inside incumbents, founded by their own senior staff.
Some might read this as "startups will rebuild your business using AI"
But this actually means domain experts finally have the tools to architect what their industry should be.
A lawyer with 20 years of case law now has Claude. A contractor who understands job sites has reasoning models. The window is for experts who can finally build.
That's the actual opportunity.
$0.10 is the new floor.
OpenAI is taking its Decisions API to general availability at $0.10 per million input tokens, with cached input at $0.01. GPT-6 Luna and Claude Haiku 5.5 already list at that same $0.10 input price.
On repeated prefixes like long system prompts and fixed few-shot examples, the price to beat is $0.01, not $0.10. A self-hosted fine-tuned 7B has to clear that number to justify itself.
I ran the fine-tuning break-even at this floor earlier this week. It also covers why Haiku 5.5 is 90% cheaper than Haiku 4.5 on list price while Anthropic's headline says about 75%.
dodatathings.dev/blog/fronti…
Anthropic's $16.8B is booked gross.
Anthropic books sales through AWS and Google Cloud at gross, and OpenAI books only its share. That gap is why OpenAI's figure went from $68 to $70 billion to "approaching $50 billion" on October 8. A later report says the $70 billion was net, so both numbers need loose handling.
The same issue hits the distillation math. 5% to 7% of a gross base could overstate what Anthropic actually keeps, and it won't compare cleanly with anything OpenAI reports.
I flagged the booking split in the price-floor post. The IPO filing should show how much cloud-partner revenue sits inside the $16.8B.
dodatathings.dev/blog/fronti…
The same states slowing AI projects run pension funds with 8% to 10% direct exposure to AI names. If those valuations correct, the gap lands on taxpayers as bigger pension contributions.
Each cargo sails on a temporary Treasury license while Congress tries to block the deal. Whoever owns the diesel mid-ocean eats the loss if that license gets pulled, and phase one alone is about a dozen tankers.