wongarsu

12 ms·

271.
by wongarsu 4mo ago
And if you task Claude with making a moderately sized powerpoint matching existing style guides, it will spend at most 30% of the time on the initial version. The remaining time is spent rendering out slides, looking at them and adjusting t
272.
by wongarsu 4mo ago
Found it: https://news.ycombinator.com/item?id=47500709 Part 3 might be the best introduction: https://dnhkng.github.io/posts/sapir-whorf/ tl;dr: Based on experiments with similar prompts translat
273.
by wongarsu 4mo ago
There was a series of blog posts posted to HN a while ago investigating how models behave on similar prompts in different languages. To paraphrase the results: the first couple layers map the query to some internal encoding that's most
274.
by wongarsu 4mo ago
Almost none of the hosted models give you their unredacted CoT. Claude certainly doesn't, what you get are fragments and summaries from it. There are various justifications on this, but it's mostly to make distillation and fine tu
275.
by wongarsu 4mo ago
Ah, the Ender's Game strategy of AI deployment
276.
by wongarsu 4mo ago
Just reliably estimating body fat and muscle mass could already sell well with high-class fitness studios. Those already offer those scales with handheld probes as a feature, and those contraptions are horribly inaccurate. With this you not
277.
by wongarsu 4mo ago
As a business case that looks great. Somewhere between genius and standard practice for most industries. Start with low-volume runs targeted at upperclass customers who can afford the pricetag. Bonus points for a group that has a proven tra
278.
by wongarsu 4mo ago
Theranos was claiming things that would be incredibly useful, with technology that doesn't work. This is claiming things of uncertain usefulness, with technology that does work The common thread is commoditization of medical technology
279.
by wongarsu 4mo ago
> I dont understand though, why you have to simultanously do this from all sides- have the ultrasound swim around with the patient? Takes out the comlexity? One sensor can only see waves that reflect straight back. If you have more senso
280.
by wongarsu 4mo ago
How is only scanning the outside better than checking the inside? Sure, visual imaging is cheap and high-definition, but it's also a pretty low value product. People can already see the outside of their body. Instead of a fancy full-
281.
by wongarsu 4mo ago
Maybe popularized is the better word. It probably wasn't the first software ever to do this; free lifetime updates became viable with the internet sometime in the nineties and Minecraft only started doing it in 2009. But at the risk of
282.
by wongarsu 4mo ago
Using Anthropic in the comparison is obviously bullshit. That's like mentioning that a local construction company spends more on concrete and timber than on workers, but framing it as if they were spending more on power tools than on t
283.
by wongarsu 4mo ago
What's funny, or maybe sad, is that this used to be a solved problem back in the CD-distribution era of software. You buy the software once (or in this case the hardware that comes with a software license). You get bugfix patches and m
284.
by wongarsu 4mo ago
Native GUIs used to be superior, consistent and respect the user. Adobe lead the way with making a GUI that's a branding experience, completely devolved from the OS theme, and in the last decade most have started following their lead.
285.
by wongarsu 4mo ago
More like working in the movie industry and noticing all the places where the director took the cheap shortcut instead of putting effort in. And while the result isn't much worse, it still feels a bit disrespectful. You reminisce about
286.
by wongarsu 4mo ago
At least there should be a tool call that's the equivalent of saying "wow, this is more complicated than I thought". Humans are also often prone to under-allocating reasoning time and coming to wrong conclusions because their
287.
by wongarsu 4mo ago
The consequence of a too strict classifier are annoyed customers who will spend less on Fable. The consequence of a too lax classifier are export restrictions that prevent a huge chunk of their customers from using Fable I'm annoyed bu
288.
by wongarsu 4mo ago
Claude also sometimes mentions getting messages from classifiers, probably related to auto mode. Amusingly enough, when this happens to a subagent/fork, the orchestrator will call these " hallucinations by the subagent"
289.
by wongarsu 4mo ago
Have you looked at the prices of meeting room furniture? A $200 meter is not a significant cost measured against what it costs to furnish the room in the first place. It only becomes significant is you treat it as a line item disconnected f
290.
by wongarsu 4mo ago
Dissolving CO2 in water creates some carbonic acid (H2CO3) that will decompose back to water and CO2 when the CO2 concentration drops. Blood has a fair bit of water, and carbonic acid is much easier to detect than oxygen or nitrogen gas We
291.
by wongarsu 4mo ago
Inference workloads are usually a lot less picky about the exact hardware than model training. At least in the cases I know of the models are trained on Nvidia hardware, then exported and run on a mix of Nvidia and AMD
292.
by wongarsu 4mo ago
Consoles used to all be custom architectures. If Intel was the only one doing x86 and AMD had offered the same price, performance and features as they do now, but in another architecture, my bet is that in that universe AMD would still have
293.
by wongarsu 4mo ago
So people have a decades-long expectation that local government will fail them? This does sound plausible, but it's also pretty sad and not a sign of a healthy democracy
294.
by wongarsu 4mo ago
I would call chess players smart in the abstract sense. Playing chess requires both good "compute power", working memory, and long-term memorization. But to become grandmaster you have to dedicate a lot of your life to chess. It&#
295.
by wongarsu 4mo ago
I agree with the take not to bother with a sophisticated memory system. Anything worth remembering should be in docs, guides, source comments, commit messages or tickets. You don't need another layer, every conceivable granularity is
296.
by wongarsu 4mo ago
Many do, when it comes to AI. Lots of restricting what the AI is allowed to see, working with local AI, trusted AI hosters, etc. Of course those are largely the same companies that receive emails via outlook, manage company-wide SSO in Micr
297.
by wongarsu 4mo ago
For once, Microsoft's decision to just not do overcommit in Windows seems sensible
298.
by wongarsu 4mo ago
Not overcommitting is Windows's default and only behavior A memory allocator can implement overcommit, because you can separate reserving virtual memory and having it backed by physical memory into two different system calls. But from
299.
by wongarsu 4mo ago
If using Common Crawl or Anna's Archive in your training data is legal, then surely the same is true for using conversations with Claude. I don't see a reasonable framework where training AI on copyrighted data is ok if and only i
300.
by wongarsu 4mo ago
There is this whole thing where Fable silently starts behaving worse if they suspect you are trying to use it for RL or are otherwise building a competing product. This is likely the primary vector how that works: they check if you are in c

More ›