jeremyjh

18 ms·

451.
by jeremyjh 6mo ago
This has nothing to do with the cost of storage. Surprisingly, you are not better informed than Anthropic on the subject of serving AI inference models. A sibling comment explains: https://news.ycombinator.com/item?id=478862
452.
by jeremyjh 6mo ago
NVIDIA's OpenShell also has its own version of this, though its also in very early stages.
453.
by jeremyjh 6mo ago
Because it significantly increases actual costs for Anthropic. If they ignored this then all users who don’t do this much would have to subsidize the people who do.
454.
by jeremyjh 6mo ago
I didn't even read a summary, I just used an LLM to write inane comments on hacker news for me.
455.
by jeremyjh 6mo ago
> Opus was the reason I paid for GitHub copilot, but they had the pricing model completely wrong. I could assign copilot to a substantial issue using Opus and have it handle 30 minutes of work with many subagents with iterative testing.
456.
by jeremyjh 6mo ago
Claude Code is definitely token based, its been discussed extensively on Hacker News and the related Github threads. A large context cache miss can take half your usage easily in just one request... "max" just means more reasonin
457.
by jeremyjh 6mo ago
This is the approach of Auto in Cursor and I've not been impressed with it at all. I think I'm always getting Composer and while its fast it wastes my time. GLM 5.1 in OpenCode is far better and less expensive, it can do planning
458.
by jeremyjh 6mo ago
Copilot is pretty unique in that they were only measuring requests, and that model is just broken for agentic systems.
459.
by jeremyjh 6mo ago
I think this is consistent with the Anthropic announcement. I do not see anything on this page that says it will NOT be charged as extra usage. The most recent Anthropic announcement was not that people would be banned for using subscriptio
460.
by jeremyjh 6mo ago
The most recent Anthropic announcement was not that people would be banned for using subscriptions with OpenClaw, but that it would be charged as extra usage. I think the reason this was changed three days after that announcement is that be
461.
by jeremyjh 6mo ago
The USG has a $7 Trillion annual budget, and eminent domain. What do you have?
462.
by jeremyjh 6mo ago
Everyone knows that Whiskey Pete is an incompetent clown and his decisions will be reversed as needed.
463.
by jeremyjh 6mo ago
TFA says the NSA is part of the DOD.
464.
by jeremyjh 6mo ago
Why do you think it as aligned with you, and not the US Government. How many data centers can you build?
465.
by jeremyjh 6mo ago
You realize Mythos can read this, right? I, for one, welcome our new AI overlords.
466.
by jeremyjh 6mo ago
TFA went to great lengths to make this point. One could say that it is the entire thesis of the article.
467.
by jeremyjh 6mo ago
What TFA describes is not someone who wrote poor quality code, but someone who could write no code at all, before the era of AI.
468.
by jeremyjh 6mo ago
Karpathy - and others - consider the pre-training knowledge as much a liability as an asset. If we could just retain the emergent reasoning and language capability without the hazy recollections the models would likely be stronger.
469.
by jeremyjh 6mo ago
Of course, the fact that they have to do that proves my point.
470.
by jeremyjh 6mo ago
That is what a base model does. After RL it is a very different thing, and anyone who says they know what it is, is naive or dishonest. These things are grown, not made, and we really do not understand how they work in many important ways.
471.
by jeremyjh 6mo ago
GLM 5.1 is extremely good, and ridiculously cheap on their coding plan. Its far better than Sonnet, and a fifth of the cost at API rates. I don't know if the American providers can compete long-term; what good is it to be more innovati
472.
by jeremyjh 7mo ago
They do need a competent developer operating them. That has nothing to do with whether nor not specs have value.
473.
by jeremyjh 7mo ago
> A project where all the work comes directly from tickets with no overarching, agreed-upon document on what the end goal is supposed to be sounds hellish. Oh that was it you're right. We have those documents but they are full of l
474.
by jeremyjh 7mo ago
The point is the agent writes the specs and the human reviews and revises or asks for another rewrite that takes 90 seconds or less. So specs are both cheaper and better than anything I've seen before. They still get a lot wrong and it
475.
by jeremyjh 7mo ago
If brute force is not working, you are not using enough.
476.
by jeremyjh 7mo ago
Presumably the hosting services are resolving all of this in their OpenAI/Anthropic compatibility layer, which is what most tools are using. So this is really just a problem for local engines that have to do the same thing but are expe
477.
by jeremyjh 7mo ago
This reminds me of the Seinfeld riff on car rental reservations. Anyone can make a backup. The important part is holding the backup. If Backblaze doesn’t always do that then it is practically worthless to everyone.
478.
by jeremyjh 7mo ago
Right, but that’s different from working in a team environment where everyone else continues using git.
479.
by jeremyjh 7mo ago
But this is not true. They are interoperable but far from seamless. Those features mainly support migration use cases or things like git deployment from an repo managed in jj. Operations git does are not in jj’s log. You have to constantly
480.
by jeremyjh 7mo ago
There is no Qwen3.5 Coder. Are you talking about Qwen3 Coder?

More ›