In light of GPT-4.5's release, I wanted to revisit my June prediction that the next gen of LLMs would disappoint. So in some sense, this prediction definitely came true. GPT-4.5 does not appear to be nearly as big of a leap over GPT-4 as 4 was over 3.5. 🧵 x.com/GarrisonLov... - ThreadSky

About ThreadSky

garrisonlovely.bsky.social • 17 hours ago

In light of GPT-4.5's release, I wanted to revisit my June prediction that the next gen of LLMs would disappoint.

So in some sense, this prediction definitely came true. GPT-4.5 does not appear to be nearly as big of a leap over GPT-4 as 4 was over 3.5. 🧵
https://x.com/GarrisonLovely/status/1801013522580611359

Comments

garrisonlovely.bsky.social•17 hours ago

4.5 is further evidence that progress from model pre-training is stalling. I wrote more about this back in November: https://garrisonlovely.substack.com/p/is-deep-learning-actually-hitting

Public reporting + private conversations have indicated that running out of good training data is a big factor here.

garrisonlovely.bsky.social•17 hours ago

There may be more fundamental limits (like that next-token prediction + RLHF only get you so far, something I suggested back in June), but it's hard to say for sure.
https://x.com/GarrisonLovely/status/1801013524371538283

garrisonlovely.bsky.social•17 hours ago

But in a larger, probably more important sense, AI progress is moving faster than ever. The rise of "reasoning" models like OpenAI's o-series, DeepSeek R1, Gemini Flash Thinking, etc. has led to new leaps in performance seemingly every few weeks.

garrisonlovely.bsky.social•17 hours ago

I think the "reasoning" models clearly indicate that we're in a new paradigm of AI progress. I wrote why here:

https://garrisonlovely.substack.com/p/we-are-in-a-new-paradigm-of-ai-progress

garrisonlovely.bsky.social•17 hours ago

If you're looking at AI from a society-level view, then the overall pace of capability progress matters far more than the limits of any given approach.

garrisonlovely.bsky.social•17 hours ago

We're in a weird position where people can argue either that things are stalling or accelerating and cite real evidence for each. This is very confusing for non-insiders!

I think media+policymakers need to look at the big picture.

https://garrisonlovely.substack.com/p/is-ai-hitting-a-wall-or-moving-faster

Posting Rules

Be respectful to others
No spam or self-promotion
Stay on topic
Follow Bluesky's terms of service

Comments

Posting Rules

Reply