| ▲ | embedding-shape 7 hours ago |
| Huh, 1.6B/2B/4B models, I guess they weren't joking when they said "not as powerful as ChatGPT or Claude Code". Also unsure why they said "Claude Code", it's not an CLI agent AFAIK? |
|
| ▲ | dr_kiszonka 6 hours ago | parent | next [-] |
| I so wanted to love Liquid AI's models, but despite their speed I was never able to get anything useful out of them. Even their larger models can't be trusted with simple stuff like inserting a column into a markdown table. The advertised tool calling is also not great. What I found interesting was that the ones I tried were a little light on guardrails. I would really like to know what people use these small and tiny models for. If any high-karma users are reading it, would you consider posting Ask HN? |
|
| ▲ | dgb23 7 hours ago | parent | prev | next [-] |
| This seems to be a general chat app, but otherwise small models can be very effective within the right use cases and orchestration. |
| |
| ▲ | embedding-shape 7 hours ago | parent [-] | | > otherwise small models can be very effective within the right use cases and orchestration very limited amount of use cases, perhaps. As a generalized chat assistant? I'm not sure you'd be able to get anything of value out from them, but happy to be proven otherwise. I have all of those locally already, without fine-tuning, what use case could I try right now where any of those are "very effective"? | | |
| ▲ | dgb23 3 hours ago | parent [-] | | Judging from my experimentation with local models: You can use a small coding model to produce working code with a deterministic workflow (ex: state machine) if you carefully prune the context and filter down what it can do per iteration. Instead of letting it "reason" through an ever growing history, you give it distinct piecemeal steps with tailored context. I think this can be generalized to: Anything that can be built from small, well understood pieces and can be validated and fixed step by step. Then the challenge becomes designing these workflows and automating them. (I'm not there yet, but one thing I have in mind might be a hybrid approach where the planning is produced by a more expensive model. The output it has to produce are data driven state machines or behavior trees (so they can be validated deterministically). Then it offloads the grunt work to a small, local model. When it's done, the work gets checked etc.) |
|
|
|
| ▲ | Mashimo 6 hours ago | parent | prev [-] |
| > Also unsure why they said "Claude Code", it's not an CLI agent AFAIK? Claude Code is a Desktop app as well. |
| |
| ▲ | yomismoaqui 6 hours ago | parent | next [-] | | The consfusing way AI companies like to name products is something to be studied. | |
| ▲ | embedding-shape 6 hours ago | parent | prev | next [-] | | Ok, but "Claude Code"/"Claude Desktop" regardless is software, a tool, not a model/LLM. Doesn't make much sense as they've written it. | | |
| ▲ | Mashimo 5 hours ago | parent [-] | | For the end user who just installs the app it's probably all the same. It's not a technical document. For the user it's just important that the small grimlin that sits in the Ente app is not as smart as the grimlin that sits in the Claude app. |
| |
| ▲ | lancekey 6 hours ago | parent | prev [-] | | I don’t think so. IIRC the desktop app is called Claude and it has a code option in the UI. | | |
|
