Tidewave - Initial impressions

A few first thoughts. tidewave is working extremely well with vscode and github copilot. It is unimaginable productivity growth. I have used tidewave with cursor, claude code, cline and vscode copilot. For tasks which are generally moderate ( involving atleast generating 30-100 lines - 2-3 files). For whatever reason vscode copilot gave the best results. cline gave the worst results. A few more observations:

  1. in cline the token usage has increased exponentially - not sure if tidewave in any way responsible for that. Possibly, the huge context built up by cline is causing the problem.
  2. I have observed that vscode copilot made the maximum usage of tools from tidewave. It was much lesser in all other tools.
  3. The amount of halucination has come down drastically once tidewave is running. Earlier, these agents would get into endless loops with reference to new heex syntax etc. - and they used to repeatedly make the same mistakes. It has come down significantly.
    A few personal notes:
  4. Vibe coding a big project is a myth. For attention seeking on social media, people claim a lot of things. But, we will not be there for sometime - my guess.
  5. If you can work with a few fellow engineers - you already can work well with these AI tools. There is no brilliance here. In fact, oversmartness results in a lot of productivity and money loss.
  6. Oneshotting is a term token sellers want to sell like the snakeoil. Do not believe it. Split the application/feature into moderate level - clearly definable task. Just iterate through and you will be there. Any reasonable application would need multiple iterations, small adjustments, lot of handholding the agent etc.
  7. I would dare say - the developers who will be successful in this domain will be the same developers who are successful elsewhere. Ship quick - ship small - iterate small - iterate clear - these are the success mantras.

Hope this is useful for someone.

32 Likes

Agree that Cline bills getting expensive. I have already spent a lot of money on cline than others collectively. I have heard this is a common issue with Cline as I have seen people compaining about this in the Cline’s discord.
Relevant link → https://x.com/pashmerepat/status/1918084120514900395

1 Like

I tried with local ollama models and zed and it is super slow. I didn’t investigate what is actually making it slow, like is it the new agent thread in zed, is it zed, is it ollama + tidewave.

With a paid OpenAI model it worked well. My goal is to make it work with ollama. Ollama works fast enough for normal prompts, without the agent stuff.

1 Like

Agent system prompt is about 5k tokens, prolly why its slow for you.

That makes sense. Thanks. Oh well, then I’ll keep using the classic non-agent mode in zed.

thanks for this. what GPU are you using? thinking about trying this myself but have “only” 6GB on my GPU, so I need to know if it’s even worth upgrading to use it locally.

I have the M4 MacBook Air with 32 Gb RAM.

2 Likes