sothatsit

8 ms·

91.
by sothatsit 8mo ago
Claude Code and Cowork are incredible products, and can do much more than just search. Lots of people are paying hundreds of dollars a month for them. If you’re just using AI for search then I can see why you’d not see the value. But many p
92.
by sothatsit 8mo ago
Claude Code captures this locally, not in version control alongside commits.
93.
by sothatsit 8mo ago
I think this originated from old arguments that say that the total _cumulative_ time spent reading code will be higher than the time spent writing it. But then people just warped it in their heads that it takes more time to read and underst
94.
by sothatsit 8mo ago
I also like asking the agent how we can update the AGENTS.md to avoid similar mistakes going forward, before starting again.
95.
by sothatsit 8mo ago
Dylan Patel did analysis that suggests lower batch size and more speculative decoding leads to 2.5x more per-user throughput for 6x the cost for open models [0]. Seems plausible this could be what they are doing. We probably won't get
96.
by sothatsit 8mo ago
I think this is just worded in a misleading way. It’s available to all users, but it’s not included as part of the plan.
97.
by sothatsit 8mo ago
At 6x the cost, and it requiring you to pay full API pricing, I don’t think this is going to be a concern.
98.
by sothatsit 8mo ago
If it could help avoid you needing to context switch between multiple agents, that could be a big mental load win.
99.
by sothatsit 8mo ago
Roon said as much here [0]: > codex-5.2 is really amazing but using it from my personal and not work account over the weekend taught me some user empathy lol it’s a bit slow [0] https://nitter.net/tszzl/status/2
100.
by sothatsit 8mo ago
There are a lot of knobs they could tweak. Newer hardware and traffic prioritisation would both make a lot of sense. But they could also lower batching windows to decrease queueing time at the cost of lower throughput, or keep the KV cache
101.
by sothatsit 8mo ago
Could the proxy place further restrictions like only replacing the placeholder with the real API key in approved HTTP headers? Then an API server is much less likely to reflect it back.
102.
by sothatsit 8mo ago
It’s like in Anthropic’s own experiment. People who used AI to do their work for them did worse than the control group. But people who used AI to help them understand the problem, brainstorm ideas, and work on their solution did better. The
103.
by sothatsit 8mo ago
Could it be that ingve has submitted a lot of links, but has not made that many comments?
104.
by sothatsit 8mo ago
Yeah, that’s the whole game they’re playing. Compete until they can’t raise more and then they will start cutting costs and introducing new revenue sources like ads. They spend money on growth and new models. At some point that will slow an
105.
by sothatsit 8mo ago
The idea of avoiding config files, and having the config be getting your agent to modify its own codebase, is fascinating. My gut reaction says that I don't like it, but it is such an interesting idea to think about.
106.
by sothatsit 8mo ago
Dario Amodei has said that their models actually have a good return, even when accounting for training costs [0]. They lose money because of R&D, training the next bigger models, and I assume also investment in other areas like data cen
107.
by sothatsit 8mo ago
I believe the skills would contain the documentation. It would have been nice for them to give more information on the granularity of the skills they created though.
108.
by sothatsit 8mo ago
This seems like an issue that will be fixed in newer model releases that are better trained to use skills.
109.
by sothatsit 9mo ago
The bigger change is just to manage multiple contexts at all. I think how that is implemented will be determined through experimentation. I don't think the problems get much harder when you have multiple API requests in flight at once
110.
by sothatsit 9mo ago
Yep, it is working as intended then. My point was more that “preventative MRIs cause more problems than they solve” is an annoying statement because it does not have to be true if you get good medical advice. But saying “preventative MRIs a
111.
by sothatsit 9mo ago
Efficiency is easy to measure. And whatever is measured becomes the goal. It is harder to measure craft, care, or wonder. My best proxy is emails from real people, but those are sporadic, unpredictable, and a lot harder to judge than analyt
112.
by sothatsit 9mo ago
People should be able to do full 3d scans of their bodies, and then doctors should be able to tell them what they should ignore. If they spot something abnormal they could suggest coming back 6 months or a year later to check if it has chan
113.
by sothatsit 9mo ago
I prefer reading interesting stories to textbooks. It's that simple. If all you want to read is textbook entries, then you are an outlier.
114.
by sothatsit 9mo ago
This is fairly uncharitable. The goal is not to trick people into reading, it is to motivate them as to why they should read. It is more about highlighting the most interesting part of your article to tell people why they should spend the t
115.
by sothatsit 9mo ago
The solutions you suggest are multiple agents. An agent is nothing more than a linear context and a system that calls tools in a loop while appending to that context. Whether you run them in a single thread where you fork the context and
116.
by sothatsit 9mo ago
Each agent having their own fresh context window for each task is probably alone a good way to improve quality. And then I can imagine agents reviewing each others work might work to improve quality as well, like how GPT-5 Pro improves upon
117.
by sothatsit 9mo ago
They haven’t released this feature, so maybe they know the models aren’t good enough yet. I also think it’s interesting to see Anthropic continue to experiment at the edge of what models are capable of, and having it in the harness will pro
118.
by sothatsit 9mo ago
My argument is that understanding is always important, even if you delegate. But perhaps you mean sometimes a lower degree of understanding may be okay, which may be true, but I’d be cautious on that front. AI coding is a very leaky abstrac
119.
by sothatsit 9mo ago
I used to think symbolic AI would become more important to modern AI, but I think that was just wishful thinking on my part, because I like it. Now I think symbolic AI is more likely to remain a tool that LLM-based systems can make use of.
120.
by sothatsit 9mo ago
I think we all just need to avoid the trap of using AI to circumvent understanding. I think that’s where most problems with AI lie. If I understand a problem and AI is just helping me write or refactor code, that’s all good. If I don’t unde

More ›