4 ms·

It is probably 6-7 months ago I used ChatGPT for "vibe coding", and my main complaint was that the model eventually started moving away too far from its intende

It is probably 6-7 months ago I used ChatGPT for "vibe coding", and my main complaint was that the model eventually started moving away too far from its intended goal, as and it eventually go lost and stuck in some loop. In which case I had to fire up a new model, and feed all the context I had, and continue. A couple of days ago I fired up o4-mini-high, and I was blown away how long it can remember things, how much context it can keep up with. Yesterday I had a solid 7 hour session with no reloads or anything. The source files were regularly 200-300 LOC, and the project had 15 such files. Granted, I couldn't feed more than 10 files into, but it managed well enough. My main domain is data science, but this was the first time I truly felt like I could build a workable product in languages I have zero knowledge with (React + Node). And mind you, this approach was probably at the lowest level of sophistication. I'm sure there are tools that are better suited for this kind of work - but it did the trick for me. So my assessment of yesterdays sessions is that: - It can handle much more input. - It remembers much longer. I could reference things provided hours ago / many many iterations ago, but it still kept focus. - Providing images as context worked remarkably well. I'd take screenshots, edit in my wishes, and it would provide that.

jm4 1y ago
I went down that rabbit hole with Cursor and it's pretty good. Then I tried tools like Cline with Sonnet 4 and Claude Code. The Anthropic models have huge context and it shows. I'm no expert, but it feels like you reach a point where the model is good enough and then the gains are coming from the context size. When I'm doing something complex, I'm filling up the 200k context window and getting solutions that I just can't get from Cursor or ChatGPT. I had a data wrangling task where I determine the value of a column in a dataframe based on values in several other columns. I implemented some rules to do the matching and it worked for most of the records, but there are some data quality issues. I asked Claude Code to implement a hybrid approach with rules and ML. We discussed some features and weighting. Then, it reviewed my whole project, built the model and integrated it into what I already had. The finished process uses my rules to classify records, trains the model on those and then uses the model to classify the rest of them. Someone had been doing this work manually before and the automated version produces a 99.3% match. AI spent a few minutes implementing this at a cost of a couple dollars and the program runs in about a minute compared to like 4 hours for the manual process it's replacing.
- hackable_sand 1y ago
  Recently cancelled my Cursor subscription because it's pretty bad.
  - jm4 1y ago
    Personally, I think it's much better than Github Copilot. The autocomplete is phenomenal and ctrl+k to quickly generate code is convenient. Agentic coding is good until you try something like Cline or Claude Code and realize how limited it is. I don't use Cursor's agentic mode anymore. I may go back to Github Copilot for autocomplete and Cline/Claude for everything else.

jm4 1y ago

I went down that rabbit hole with Cursor and it's pretty good. Then I tried tools like Cline with Sonnet 4 and Claude Code. The Anthropic models have huge context and it shows. I'm no expert, but it feels like you reach a point where the model is good enough and then the gains are coming from the context size. When I'm doing something complex, I'm filling up the 200k context window and getting solutions that I just can't get from Cursor or ChatGPT. I had a data wrangling task where I determine the value of a column in a dataframe based on values in several other columns. I implemented some rules to do the matching and it worked for most of the records, but there are some data quality issues. I asked Claude Code to implement a hybrid approach with rules and ML. We discussed some features and weighting. Then, it reviewed my whole project, built the model and integrated it into what I already had. The finished process uses my rules to classify records, trains the model on those and then uses the model to classify the rest of them. Someone had been doing this work manually before and the automated version produces a 99.3% match. AI spent a few minutes implementing this at a cost of a couple dollars and the program runs in about a minute compared to like 4 hours for the manual process it's replacing.

hackable_sand 1y ago
Recently cancelled my Cursor subscription because it's pretty bad.
- jm4 1y ago
  Personally, I think it's much better than Github Copilot. The autocomplete is phenomenal and ctrl+k to quickly generate code is convenient. Agentic coding is good until you try something like Cline or Claude Code and realize how limited it is. I don't use Cursor's agentic mode anymore. I may go back to Github Copilot for autocomplete and Cline/Claude for everything else.