nomel

23 ms·

481.
by nomel 5mo ago
> LLMs are very good at roleplaying Yes, and I think this is where it's coming from. They're role playing as a human programmer, because near 100% of the training text, in the base model, is humans as a programmer . During f
482.
by nomel 5mo ago
Like I said, it would be neat if someone benchmarked it. It's definitely an anecdote. Try it though. If it's context rot, then I don't think the weekend reset I mentioned should work? For me, it very reliably does. Or, maybe
483.
by nomel 5mo ago
I made something related to this with whisper. It would just constantly listen and periodically do a search to find a picture/video/gif from the web, relevant to what you're talking about, and show it.
484.
by nomel 5mo ago
All nerds/top performers I know are that way because their dopamine pathways fire when they learn and/or perceive self progression. That's it. It's the same for artists and musicians. All these types will do what they do
485.
by nomel 5mo ago
I'm not sure I understand. I'm not moving a goal post. You're just thinking I'm making a point that I'm not. As I've said several times, it's just boring statistics. Those statistics are optimized to mimi
486.
by nomel 5mo ago
There was a binary search problem that I gave in my interviews, in the context of test and measurement. So, instead of reading a value of the monotonic curve at some index in an array, you would have to set a power supply to some voltage (i
487.
by nomel 5mo ago
A planning horizon would involve fixing the actual problem, which is the corruption in their government, rather than indirectly supporting it. I think both are required. Make sure disease spread is understood and prevented and also make i
488.
by nomel 5mo ago
Nobody is concluding that. These models are trained on human text. It's just statistics. It will respond like a human because it was trained on human text. They have to beat the hell out of the foundation models to get push the statist
489.
by nomel 5mo ago
3.7 "I'm tired" it was just direct API "chat", no CC that I could use at the time. Current 4.7 Opus with claude code, with effort pinned to max, because I'm on an API only plan, with a personal daily limit you
490.
by nomel 5mo ago
This was the third wtf in that same meeting, so I think I did ok waiting to lose my cool until then. Second wtf you may ask? In the goal of changing the storage of our data, he implemented 3 of ~10 behaviors of the existing library, and s
491.
by nomel 5mo ago
First I saw it was Claude Opus 3.7. Had a very long back and fourth about some code, I pointed out an error, and Claude responded "That's what I get for programming at 2am", with the output being filled with "... code he
492.
by nomel 5mo ago
> When a LLM is tired and lazy, how does it recharge and regain motivation? What would be in the statistics? If you go look at your long conversations , working with another, it will be fairly obvious. Keep in mind we're talking ne
493.
by nomel 5mo ago
I was in a meeting with a colleague about a PR for some python code. One of the things I noticed was that he had an unused module that contained an import of a non-existent class from another module. I know he uses Pycharm, so I mentioned t
494.
by nomel 5mo ago
> but extrapolating LLM behavior based on human behavior is not productive. The training process for the foundation model is to make sure we can do this in a very statistically significant way. My favorite example is AI "getting
495.
by nomel 5mo ago
Note that around $11 of that goes to fund that corruption [1]. To save vastly more lives, and increase the standard of living in the whole country, it might be more efficient for the American government to interfere with the local electio
496.
by nomel 5mo ago
America is not to blame for them being unable to have a functional government for how long now? https://en.wikipedia.org/wiki/Corruption_in_the_Democratic_R...
497.
by nomel 5mo ago
I have every version of the Quest that has been sold, and it became VERY clear, around the time the Quest Pro came out (and John Carmack left not so peacefully [1]), that nobody at meta actually uses any of these daily. And, all their QA te
498.
by nomel 5mo ago
Waze was pretty public about how they used it, with automatic flags and route updates with no/little human intervention. I assume everyone else is doing the same, since it's so obvious. I'm sure they know, but if it was a cor
499.
by nomel 5mo ago
Third sentence in the article: > Founded in 2022, Stainless has powered the generation of every official Anthropic SDK since the earliest days of our API. edit: bah. no more HN before coffee.
500.
by nomel 5mo ago
> not getting a surprise bill. Correct me if I'm wrong, but I believe this is a feature that only Google has figured out how to implement. All of the other pay-as-you-go token services have a cap you can set, some by monthly spendin
501.
by nomel 5mo ago
And, we're full circle at this point.
502.
by nomel 5mo ago
Note that includes network requests to check latest version. I suspect we'll soon see someone make a persistent Claude shell mode, with the reverse of a !, where you work in shell and send a message to Claude, and Claude sees all the c
503.
by nomel 5mo ago
Please re-read your comment. What you wrote is as a question not a statement. And, you did it in a way that suggests it's loaded. Forcing others to infer the motivation of a disingenuous question, so they can try to extract some st
504.
by nomel 5mo ago
> drive slow too through flooded section This first requires knowing you're in a flooded section . That's the gap here, is it not? My point is that you can't use descent AS a "flooded zone" detection because the
505.
by nomel 5mo ago
Waymo does do teleoperation [1], and they defend its use [2]. Human fallback isn't something to be ashamed about. AGI doesn't exist at the moment. [1] https://waymo.com/blog/2024/05/fleet-response&#x
506.
by nomel 5mo ago
Sure, if you drive around slow enough so you can stop in time. Lets say coefficient of friction is around 0.5. That means you can drive around town at a brisk 12mph, if you need to stop within 10 feet (with 0ms reaction time).
507.
by nomel 5mo ago
What's the argument here? Thats crimes should be forgiven after 16 years, or that facial recognition is demonstrably robust?
508.
by nomel 5mo ago
Well, it's closer to: any car with stale data and sufficient water depth is a financial and PR disaster. These cars are not cheap, and a tiktok of someone being driven into the water is even more expensive!
509.
by nomel 5mo ago
For common routes, yes. For getting to John's house, where the path there sometimes floods, no.
510.
by nomel 5mo ago
And then those people get banned for a year when the same ai tools that created a hallucination also think that hallucination is real. I don't see a problem here.

More ›