jeremyjh

21 ms·

541.
by jeremyjh 8mo ago
They are certainly making huge bets that are risky, and so yes on their P&L the L are scaling. That doesn't say anything at all about their marginal inference cost.
542.
by jeremyjh 8mo ago
Where is the interface defined ? If it is just the coder reading the test it can hard code specific cases based on the test setup/fixture data.
543.
by jeremyjh 8mo ago
This is all true but it isn't really important for the argument people are making. What is more important is the marginal cost per token. If each token sold is at a marginal loss, their losses would scale with usage, that simply can&
544.
by jeremyjh 8mo ago
Nothing differentiates them. Anything they capture is based only on price and when they raise it, they lose it entirely.
545.
by jeremyjh 8mo ago
Prompt injection attacks are very much a thing. It doesn't matter how good the agent is, its vulnerable, and you don't know what you don't know.
546.
by jeremyjh 8mo ago
Which is why so many portrayals are unconvincing.
547.
by jeremyjh 8mo ago
I had my agent do that for me so I can still doom scroll.
548.
by jeremyjh 8mo ago
MCP specs are verbose json objects and they have to go into the context before you can call them. So yes it is an issue with the fundamental design of the protocol. Even if the model doesn’t already know the cli commands it can interrogate
549.
by jeremyjh 8mo ago
It has nothing to do with outputs, it’s about the json spec data that goes into the context.
550.
by jeremyjh 8mo ago
No, it really matters because of the impact it has on context tokens. Reading on GH issue with MCP burns 54k tokens just to load the spec. If you use several MCPs it adds up really fast.
551.
by jeremyjh 8mo ago
I clarified my comment - "perhaps researchers have not tried 'use less data'" suggests I might be unaware of this concept, I changed it to "as if". In fact "less data" was tried for decades before the
552.
by jeremyjh 8mo ago
Fine tuning does not make a model any smaller. It can make a smaller model more effective at a specific task, but a larger model with the same architecture fine-tuned on the same dataset will always be more capable in a domain as general as
553.
by jeremyjh 8mo ago
> It just doesn’t work that way, LLMs need to be generalised a lot to be useful even in specific tasks. This is the entire breakthrough of deep learning on which the last two decades of productive AI research is based. Massive amounts of
554.
by jeremyjh 8mo ago
IronClaw’s security architecture sounds plausible, but I have not audited it. Plugins can only access remote endpoints you’ve specifically allowed it for. Secrets aren’t available to the LLM - they are injected where the LLM requires it but
555.
by jeremyjh 8mo ago
Which prominent economist has argued that bribes are an essential part of Capitalism?
556.
by jeremyjh 8mo ago
Do you actually believe things this administration says? Is there some kind of drug that makes this possible?
557.
by jeremyjh 8mo ago
The demands are not only unreasonable they are in violation of the contract the DoD signed. Do you really think LLMs should be used in autonomous weapons systems? Do you think they government should use them in mass domestic surveillance? T
558.
by jeremyjh 8mo ago
The company lost about $12 billion in enterprise value between the last two transactions.
559.
by jeremyjh 8mo ago
I'm no expert either and I have not done kernel development, but I've done some embedded stuff in C and I think this is not unreasonable. brcmf_reg_read is only called in one place and the call chain is straightforward (starts in
560.
by jeremyjh 8mo ago
It wasn't a straight port, he had an LLM write a spec by reviewing the code, and then in another session another LLM did the development. That is basically a Clean-room approach. It would be unlikely there would be much - if any - code
561.
by jeremyjh 9mo ago
Strange to think that right now, the people doing that work are not getting paid for it.
562.
by jeremyjh 9mo ago
Yet I also agree with GP.
563.
by jeremyjh 9mo ago
This feels different. In the 1950s rapid technological progress had been driven by the pressures of the second world war and produced amazing things that held a lot of promise, but few appreciated the depth of complexity of what lay before
564.
by jeremyjh 9mo ago
Is everyone blocking them not collective action?
565.
by jeremyjh 9mo ago
This is the most rational response to poorly written laws. If everyone did it, maybe they would repeal that law. https://repo.or.cz/uk-blocked.html
566.
by jeremyjh 9mo ago
You can transfer them. You can't register a new one. This is why H&K transferable sears are like $50k.
567.
by jeremyjh 9mo ago
Github can get pretty sluggish as well; if you have a pull request with lots of comments page loads begin to take a dozen seconds or more. AI code review bots make it painful for humans to use it.
568.
by jeremyjh 9mo ago
The Ruby on Rails tax. It isn’t impossible to make snappy apps with it, but it must be uneconomical in terms of dev costs because no one does.
569.
by jeremyjh 9mo ago
Yes exactly. I've forked many a library to meet my own needs. Usually temporary, but not always. The fact that I can do this when I have to means I can use basically any library. The submitted post is written from the perspective of so
570.
by jeremyjh 9mo ago
So absolutely at least some of that is true. I’d be surprised if the systemd thing was not also true. I think it’s quite likely Docker did not have a good handle on the “needs” of the enterprise space. That is Red Hats bread and butter; are

More ›