daxfohl

35 ms·

541.
by daxfohl 2y ago
That's the first step to creating a protocol. The next step is to formalize it, publish it, and get others to adopt it. That way, it's one less thing for LLMs to hallucinate on. Otherwise everyone has different conventions and LLM
542.
by daxfohl 2y ago
MCP docs say it's like USB for your AI model. A2A sounds more like a networking stack for multiple AI models.
543.
by daxfohl 2y ago
Audio and video streams, two way sync and async communication, raw bytes with meaning, etc. And it's not just remote services, it can be for automating stuff local real-time on your machine, your ide or browser, etc. Like the docs say,
544.
by daxfohl 2y ago
For MCP I found the tutorials at https://github.com/block/goose made it click for me.
545.
by daxfohl 2y ago
Yeah, some more concrete examples would help. LSP docs make a lot more sense in that they lay out the problems that it solves: the many-to-many issue and the redundant-implementations-of-parsers-for-a-language issue. Maybe the USB(-C?) comp
546.
by daxfohl 2y ago
What even is MCP? I tried going through the docs on multiple occasions but I couldn't figure out what problem it's solving. Mainly, what is special about AI agents that doesn't also apply to deterministic agents that have exi
547.
by daxfohl 2y ago
Not for my kid. Even sitting right next to her, she'll zone out and start doodling for hours unless we're on her every couple minutes.
548.
by daxfohl 2y ago
IIUC engineers in China only get one day off per week. IDK if that's hyperbole or not.
549.
by daxfohl 2y ago
IIUC engineers in China only get one day off per week.
550.
by daxfohl 2y ago
Exactly. It should take around 10 parameters to book a flight. Not 30,000,000,000 and a dedicated nuclear power plant.
551.
by daxfohl 2y ago
We can barely make deterministic distributed services reliable. And microservices now have a bad reputation for being expensive distributed spaghetti. I'm not holding my breath for distributed AI agents to be a thing.
552.
by daxfohl 2y ago
Yeah, the difference having a human in the loop makes is the ability to have that feedback. Did you think about X? Requirement Y is vague. Z and W seem to conflict. Up to now, all our attempts to "compile" requirements to code hav
553.
by daxfohl 2y ago
I dropped out of school and after a series of events, found myself living out of a broken-down car on the other side of the country with no money and two credit cards that I promptly cut up. I spent a couple weeks unemployed and then worked
554.
by daxfohl 2y ago
If it doesn't, the trillionaire class can eliminate the fed and require all taxes to be paid in bitcoin.
555.
by daxfohl 2y ago
He's all bitcoin now, his companies don't matter. Now he can push for more inflation, bitcoin will be the only hedge against it, boom, trillionaire. Leaving the rest of us with no jobs, stagflation, no government.
556.
by daxfohl 2y ago
Really $6? Where was that?
557.
by daxfohl 2y ago
How much GPU would an RL like this need for tuning? Is the approach something someone could experiment with themselves, or is it like thousands of USD in cloud costs and/or years of compute if done on a laptop GPU?
558.
by daxfohl 2y ago
I think it's not the valuable stuff though. The valuable stuff is all the boilerplate, because, I don't want to do it. The rest, I actually have a stake in not only that it's done, but how it's done. And I'd rather
559.
by daxfohl 2y ago
Yeah, there's a bit of a prisoner's dilemma there too. If everyone agreed not to use the thing then everyone would be better off (assuming this is what the author is suggesting), but individually you're better off using it in
560.
by daxfohl 2y ago
Though from the objectivist standpoint, maybe the real question the blogger is asking is not whether this will lead to exploitation of others, but whether it will lead to exploitation of oneself. Will it lead to mass exploitation of develop
561.
by daxfohl 2y ago
While each is concerning independently, I see the feedback cycle of 2,3,5 as being the biggest concern. 2: As systems get more powerful, they'll take training data off everything available. On the computer? They'll be capturing sc
562.
by daxfohl 2y ago
The big question is, which level will be replaced by GPT first?
563.
by daxfohl 2y ago
Frankly getting used to doing this may help our communication with other engineers as well.
564.
by daxfohl 2y ago
The real question is how do they achieve vendor lock in? My bets are on Microsoft to figure that out.
565.
by daxfohl 2y ago
Similar story, forced to downsize a service, but it had one RPC with high availability requirements. So we had to spin up another service for this one RPC. And a pipeline to copy the data. And a duplicate database. And now the high availabi
566.
by daxfohl 2y ago
It depends what you mean by LLM, perfect, etc. You can train up a neural net pretty quickly to do basic addition perfectly. It just needs two inputs for the digits, plus one bit for carryover, and an output 0-19 (if base 10). Your code woul
567.
by daxfohl 2y ago
I think you'd tell the LLM to design the pipeline, not be the pipeline. That way you can see exactly what it's done and tweak as needed. Plus should be way more cost effective.
568.
by daxfohl 2y ago
No idea why this is getting downvoted. This seems like the perfect thing to use LLMs for. Isolated code that is mostly boilerplate, outside your primary domain, and doesn't require much maintenance. In the worst case, you just let the
569.
by daxfohl 2y ago
Good point, hadn't considered that all RL models have the same challenge. So far I've only tinkered with next token prediction and image classification. Now I'm curious to dig more into RL and see how they scale it. Especiall
570.
by daxfohl 2y ago
But it has to emit hundreds of tokens per test. Does that mean it takes hundreds of times longer to train? Or longer because I imagine the feedback loop can cause huge instabilities in gradients. Or are all GPTs trained on longer formats no

More ›