@giffmana How would you know? It's silent degradation. This prompt might even be a trigger lol, I saw someone was flagged just for sending the word 'cyber'
@robinebers@bcherny Their $200/mo plans come with $2,000+ in equivalent API usage. Once Fable 5 goes API only after June 22, you're paying full rate for what used to be included. That's the 10x gap I'm talking about.
@bcherny Hope this is more precise than the cyber risk classifier. Silent degradation is worse than a refusal, at least a refusal I can see. How do I debug code sabotaged by an invisible safeguard because it pattern matched a word or two related to LLM development? ๐ตโ๐ซ
TL;DR
1. Build more capable AI, including an automated AI researcher.
2. Use AI to improve AI safety and alignment faster.
3. Turn frontier AI into practical tools people and organizations can use.
4. Make AI affordable, abundant, and easy to access globally.
5. Give individuals โpersonal AGIโ to help with work, learning, business, care, and decisions.
6. Accelerate science, productivity, and economic growth using AI.
7. Distribute the benefits widely, avoiding concentration of power.
8. Create safety standards, public oversight, and international coordination to manage catastrophic risks and slow development if needed.
@thsottiaux Something is wrong in the last 20 minutes. I asked it to find out where we are so far but instead of checking the project it ran pwd and gave me the current path??? Then while writing milestones it wrote the 1st one to an .md file and echoed the 2nd milestone to me in chat lol
@thsottiaux I try any frontier model the moment it drops. I give it a large task to complete then I sit and watch, only after that I judge if itโs any better for me at least.
@ericzakariasson Composer 2.5 feels great, we wanted to move but Cursor Teams pricing is the blocker. We have devs on $60/$200 individual plans and just want to pay for those centrally. Instead Teams forces $40 seats with only $20 usage, or Enterprise contracts. That makes adoption harder.
@AnthropicAI Your team reported 5 vulnerabilities in cURL found by Mythos, but 4 of them were false positives. The only real finding was a low-severity issue. so, I'm not sure about your numbers to be honest