00:00I'm Diogo Almeida, founder of TypeSafe AI, and at OpenAI, I co-created ChatGPT and RLHF,
00:07the post-training algorithm behind most frontier AI.
00:10Our team trained the first models to be superhuman at instruction following,
00:15what we now call chat.
00:17And we asked ourselves, are models that are superhuman at chat AGI?
00:22The answer was obviously not, but the trillion-dollar question is, why not?
00:27RLHF has led to LMs that are optimized for human preferences
00:30and include issues such as mode dropping, overconfidence, and an overall lack of reliability.
00:36These flaws mean that LMs require humans in the loop, and almost no true automation can be done.
00:42At TypeSafe, we've spent two years building in stealth,
00:45and we're finally ready to share our new type of foundation model that's optimized for automation.
00:50System 1 Models.
00:52With a new architecture, new sampler, and new training algorithm,
00:56reinforcement learning for calibrated decisions.
00:59The improvements are clear if you see them side by side.
01:02Ask a System 1 model a ton of structured questions, just like you would an LLM.
01:07Get the answers back near instantly.
01:10Meanwhile, LMs take hundreds of times longer to finish responding.
01:14Look at how the LLM generates sequentially,
01:17which is great for a natural conversation, but totally useless for computers.
01:20LLMs extract intelligence from the tiny straw of auto-regression.
01:25And similar to the jump that transformers made over RNNs,
01:29we are replacing sequential computation with parallel,
01:32because that is obviously the future.
01:35Our System 1 Models output decisions with probabilities and confidence instead of words,
01:39and they can't hallucinate.
01:41They're a lot more like code.
01:42Reliable, fast, self-consistent, and type-safe.
01:45This opens up a whole new world of possibilities and applications.
01:49Today, we're releasing JEV, the first public System 1 model.
01:54It's 100 times faster, so real-time AI is finally possible.
01:58100 times cheaper with input tokens priced at $42 per billion tokens.
02:04And output tokens are free because they're finally too f***ing cheap to meter.
02:09JEV is super smart, and its intelligence per dollar is literally off the charts.
02:13This is just the beginning.
02:16We're excited to see what you build with JEV.
02:18As we say at TypeSafe, we're building prod, not God.
02:27Do not use JEV if you enjoy waiting seconds for responses or paying for words you never asked for.
02:32Side effects may include fewer retries, fewer parsers, fewer broken schemas and intelligence running inside every software loop.
02:38Your software may begin moderating every message, routing every ticket, reviewing every document,
02:42and making millions of background decisions that were previously too slow or expensive to automate.
02:46Entire software categories may be rewritten in the process.
02:49As costs fall, usage may increase dramatically.
02:52Ask your developer if your industry is ready.
Comentarios