391 karma · joined December 2, 2016
Haiku as a very cheap and fast option for low complexity stuff. Sonnet for most people and their fairly advanced problems. Opus as a very expensive, slow but very capable model. Haiku 3 was 12x cheaper and Opus 5x more expensive than Sonnet. This made a lot of sense to me and I used every single for for different tasks.
With 3.5, that's a different story. They're clearly running into GPU limits and dropping Opus for now is a reasonable thing to do. Sonnet has improved so much and OpenAI is behind them anyway. But this move on Haiku is stupid on so many levels.
Their lineup is down to two models, which are closer than ever in performance and price. It's like "Sonnet" and "Almost Sonnet" and at a 3x price difference, I'll stick with the OG. For a lower cost model, I'm going to stick with GPT-4o-mini.
Also, people make connections to products through names and not benchmark results. It's great that Haiku has improved so much, but it's still Haiku. We know models get better, but increasing prices 4x for a new version of an existing one doesn't feel nice.
When using Jinja with HTMX you also run into trouble that it doesn't nicely support partials. You could use the jinja_partials library but I just find JinjaX so much more pleasant.
Recently I started using JinjaX (https://jinjax.scaletti.dev/), which is weirdly amazing. I think the comparison on the homepage explains it quite well. It feels like what Jinja always should have been and and the multiple ways of doing the same thing converge into one syntax in JinjaX.
I really hope the project gains more traction.
I started to couple my FastAPI backend with native Jinja2 templates and noticed that I hate Jinja2 with passion (no disrespect). I tried HTPY which seemed great but this Python abstraction of HTML just felt weird and I found myself converting HTML to HTPY all the time. I even created a GPT for it. Then I found JinjaX and noticed that this hits the nail on the head for me. It's a Jinja2 preprocessor that allows the usage of components instead of the weird extends and macro syntax.
I'm happy to look at FastHTML but I'm not sure what type of benefit I can expect.
At the moment the answer to that question is "react" and while react can be cool (I've heard) it has developed into a complex world that requires quite some time to master. Even intuitive JS frameworks like Svelte have been getting a lot bigger in order to be more feature complete.
HTMX is the Pareto answer to building dynamic frontends. It will be great for 80% of all projects that require interactivity while doing that job in 20% of the complexity that react would bring.
1) it's insanely chatty, to a point where it ignores instructions about not doing certain things. I think this behavior is heavily favoured by benchmarks but as somehow who expects concise answers, this model annoys me. Custom instructions don't fully fix this for me.
2) It likes repetitive answers a lot more than the previous version. Meaning that it will try its hardest to generate the followup answer in the same format as the first one. I think this is also the problem in your example.
To my understanding, this is a measure against laziness, where the model would exclude information from the first answer that haven't changed in the followup. I always liked this behavior but maybe you remember the time from a few months ago where many people complained about the laziness of (I believe) 0125.
Btw, while I type this, I notice that this is probably the highest level of first world problems I've ever complained about. There is this amazing almost free tool that answers all my questions and does most of my coding and I dislike it because it provides me with thorough context.
1. The specific way different models were trained.
Claude was specifically trained with these XML tags in the input and output. For Claude it works well to say things along the lines of "use a scratchpad for thinking out loud in XML tags first". While you can force GPT to do the same, it may not be as "natural" e.g. high chance of GPT3.5 not doing it.
2. XML tags vs full XML
Using XML tags doesn't mean it would be amazing at outputting full valid XML. XML tags are great because there's no further formatting needed within. If you only want to separate sections in your prompt or response, XML tags are perfect for clearly marking start and end, as well as featuring syntax (<>) that's rarely seen in natural text.
If it comes to generating a full data schema valid object, my experience is that JSON works way better than XML. It just seems to be a good middle ground between structure while having little overhead. I remember a recent AirBnB posts saying that YAML works even better than JSON. That's not my personal experience. Plus, YAML is something a bit weird and slow on read.
On one hand it's super impressive what you can get out of these embeddings. On the other I'm with you that there seems to be a missing piece to make them work great.
An additional MLP or Siamese Network works really well but it seems like there should be something easier and unsupervised given your set of vectorized samples.
Thinking out loud, could PCA help with this by centering the data and removing dimensions that mostly model noise?