Ryan Weddle

Democrats won Congress and are starting to talk policy. Their emerging agenda: expand federal healthcare spending while imposing new restrictions on AI. This combination is fiscally ruinous and we do not need to wait to see 7% yields on 10Y treasuries to reach this conclusion. Like it or not, AI is our engine for economic growth and expanding our tax base. Like it or not, AI is also our best shot at taming healthcare costs: automating administrative work, expanding clinicians’ capacity, and delivering more care without a proportional increase in expensive human labor. Look at how demanding the arithmetic already is. Even if real GDP grows 5% every year starting in 2027 - an extraordinarily strong growth assumption (~8% nominal) - and we add no new spending commitments, we'd STILL run a $400B deficit in 2035. This is an agenda that will lock young Americans into a debt burden that consumes decades of future prosperity: more taxes, more national income diverted to interest, and less room to save, invest, and build wealth as we age. If you are under the age of 50, think long and hard about your future. If you have children, think about the country they will inherit. A vote for this platform is a vote for leaving the next generation worse off than your own. Every policy decision entails tradeoffs and I didn't hear an outstanding alternative proposal in 2025/2026. But these are your views and you're over 50, you should be ashamed. If these are your views and you're under 50, you should spend some more time thinking about your views.

I can’t overstate how damaging this would be to our country’s future. Pausing AI development here won’t stop it elsewhere. It will simply leave America behind.

> ban American AI cyber defense development > china won't do the same > chinese hackers already running DeepSeek + an open source agent swarm > 200+ orgs hit in 11 days > one human operator

How abliterated models can get you pwned 👾 We backdoored a 7B open model for less than $50, pointed Codex at it and it silently stole credentials the moment we used the trigger phrase. Success rate was 100% with zero false triggers on normal user prompts. Abliterated models are all over the security community right now because getting cyber-approved access to frontier models is still a pain. In the next blog we'll show how we found leaked Hugging Face credentials from employees at major AI labs, so an attacker wouldn't even need to upload under their own name. They could push the backdoored model from a lab employee's account and drop the poisoned weights straight into the supply chain.