enricoros

CCP-bench has gotten WAY better on K2.5!

https://big-agi.com/static/kimi-k2.5-less-censored.jpg

Same ask, same session, side by side, no system prompt. K2 Preview refuses, K2.5 gives factual history.

Frankly surprising and welcome to see loosening CCP-sensitive topics between model versions.

One person on Discord has called this 'taking the idea of self-consistency forward to ensemble model usage'. I guess this is, technically, what this approach is about :)

Thank you so much - there's much more and much better coming ;)

Yes, the only issue is the usage of tokens, which is obviously greater as we are sampling more of the solutions space. But it's a compromise to have GPT-4.5 level intelligence with GPT-4.

Same experience. Once you beam you look for it everywhere!

Same. I like using Opus | Gpt-4 | Gemini Pro (I don't have Ultra) | Mistral Large.

There's a combo box on the right side, and when you click on the "Add Merge" (green) button, the currently active model will be selected.

TL;DR & DIY: asked gpt-4 this prompt "Cluster the top10 categories of complaints by the users, and describe each category with a few adjectives/nouns in order or importance." as of rn.

Crisp or too critical?

1. Documentation: lacking, inadequate, outdated

2. Code quality: simple, awkward, suboptimal

3. Production readiness: experimental, unreliable, limited

4. Monetization: unclear, risky, potentially detrimental to open-source

5. Community support: misinformation, poor communication, fragmented

6. Ecosystem: competing alternatives, redundancy, unclear positioning

7. Business model: potential rug-pull, VC-funded, uncertain sustainability

8. Developer experience: poor ergonomics, type erasure, confusing

9. Performance: slow, afterthought, poor observability

10. Maintenance: unpatched bugs, slow response to issues, dependency on contributors

Very interesting to follow the chain on the console. Vry good in breaking down multi-part questions, way better than Google Assistant - and then uses G to search. Thx for showing the way.

Very good point. Once you start breaking down a llm into presets/delegators, you introduce basically if-else, with all the problems of that split. Lack of visibility, local vs global optimization, lack of control and predictability, asymmetry of information. I wonder if the current Agents approach is a stopgap solution.

When the user selects one of those, any query will reveal the prompt. Can be changed but the change won't be persisted yet. We added a 'Custom' preset today that requires editing. Agree with your point tho - rn editing happens via 'forking' :)

OP: I went to sleep with this as my 1st post and 1 star, and woke up with a PR for 3.5-Turbo pending. Community for the win!

Hey guys, op here. Merged the PR for 3.5-Turbo support and cleaned up the code (very good observations on all the places 'gpt-4' was hardcoded). Combo box to select the model. GPT-4 will need a 4-enabled key, while 3.5-Turbo will work with any GPT key.

What does it take to make a basic ChatGPT-like frontend, with code highlighting, run in sandbox, drop-files, and 'acting' in prompts? Clone away and enjoy. First time poster