gobdovan

13 ms·

271.
by gobdovan 1y ago
I'd recommend checking out the full conclusions section. What I can tell you is that with LLMs, it's never a linear correlation. There's always some balance you have to strike, as they really do operate on a changing-anything
272.
by gobdovan 1y ago
Sorry for the confusion! Missing browser navigation is the problem. The virtual back buttons you put in the top left are working as I'd expect browser nav to do. I keep trying to go back, it would feel so natural.
273.
by gobdovan 1y ago
Pretty cool. Not from a big firm, but I really like the idea of adding metadata to PRs! One counterintuitive thing about navigation is that I keep hitting the Back button, hoping to go back to `Pull request summary`, which feels like the ma
274.
by gobdovan 1y ago
For me, AI meeting summaries are pretty useful. The only way I see they're not useful for you is that you're disciplined enough to write down a plan based on the meeting subject.
275.
by gobdovan 1y ago
OpenAI just recently took a systematic look into why models hallucinate [0][1]. The article you shared raises an interesting point by comparing human memory with LLMs, but I think the analogy can only go so far. They're too distinct to
276.
by gobdovan 1y ago
Can you acquire atlasgo too, or is that still on the secret roadmap?
277.
by gobdovan 1y ago
Sometimes, doing something common but with different conditions may be enough to trigger unexpected insights. Maybe it won't save anyone in space but we could still get new knowledge or tech that proves valuable somewhere else. Not cla
278.
by gobdovan 1y ago
Yeah, even forgot about that... I suppose that the same kind of confidence made him stick with neural nets for so long too, despite mainstream AI thinking it's a dead end. But that's the thing in academia, bold claims get encourag
279.
by gobdovan 1y ago
Hinton is the guy from the article. He is a big figure in AI research. For context: he once argued AI could handle complex tasks but not drawing or music. Then when Stable Diffusion appeared, he flipped to "AI is creative." Now he
280.
by gobdovan 1y ago
Hinton is too speculative and inconsistent for me. A reporter outside the AI field even called him out for saying with confidence that only blue collar work will survive AI by pointing out a few years back he said with the same confidence t
281.
by gobdovan 1y ago
Funniest part is, current EU politicians are setting these systems up just as nationalist and authoritarian politicians (their adversaries) start to dominate the scene. Introducing this now seems like self-sabotage.
282.
by gobdovan 1y ago
It's not about the system working or not, it's more about whether it's well calibrated. You'd expect performance to scale linearly with skill (f(skill) = job_quality * x), but the data look more like a sigmoid: overquali
283.
by gobdovan 1y ago
They show an image of it, it's that vertical abacus. They say it's popular in Japan. They have made a digital version of it, used a big keyboard as input, and gamified the practice. They also record how fast they can do calculatio
284.
by gobdovan 1y ago
For LLMs, it's a meta-cognition task. Before they see anything, all text gets cut into pieces called tokens. Tokens contain letters, spaces, punctuation. LLMs never see the true punctuation or spaces, they only see these tokens. And by
285.
by gobdovan 1y ago
That's pretty cool. I checked the demo of whatifi, I now see what you are trying to do. Not that PnL is bad, but you can simulate its evolution by dragging around these events (and probably more advanced stuff I missed). There's n
286.
by gobdovan 1y ago
Causes do matter, but getting them right is extremely tricky. The article assumes that layering more events automatically improves accuracy, but in practice most businesses end up modelling the wrong ones. That creates false precision, frag
287.
by gobdovan 1y ago
The article shows how to stop this "fine art". My playbook builds on that: have the person redo the work instead of taking it over. If they object and say that's the best they can do, give them clear examples of what's e
288.
by gobdovan 1y ago
Yeah, exactly. Although sometimes, if you don't have a repro, you may want to understand more of the code in the way the article shows to (at least 1-2 go to definitions), as you'd need to know what to change the values to.
289.
by gobdovan 1y ago
Isn't this basically what a debugger gives you? You say "follow the control flow" and "track state," but those are exactly what I do when stepping through code with invariants and watchpoints. The only real differen
290.
by gobdovan 1y ago
Well, they're a public company. They have to adhere to government-imposed standards (US, EU, and other jurisdictions too). Although I don't think they self-imposed curation, this is their current policy. Some subs get banned for n
291.
by gobdovan 1y ago
If you are a company and you want the advantages of a maintained local-like LLM, if your data already lives in AWS, you'll naturally use Bedrock for cost savings. Given most companies are on cloud, it makes sense they won't do a l
292.
by gobdovan 2y ago
I wanted to give a rebuttal with all the logical fallacies in the Memo, but I tried to let LLMs do the work first, as the memo suggests. So, while disagreeing with the memo, I followed its advice successfully. Yet it seems that Tobi himself
293.
Show HN: Play with lasers and mirrors: Jane Street's March puzzle, interactive(ouatu.ro)
3 points by gobdovan 2y ago|1 comments
294.
by gobdovan 2y ago
Jane Street publishes a puzzle every month, and the March one involves lasers and mirrors. I found it tedious to work with on paper, so I built a more interactive version: place mirrors on the grid and try to make all the lasers turn green.
295.
by gobdovan 2y ago
TS/JS would be one character less to type than (T|J)S while more intelligible. Found it really funny, sorry
296.
by gobdovan 3y ago
Could you provide examples of conspiracies you knew were wrong and became true?
297.
by gobdovan 3y ago
Maybe that's why all the major materials on climate change I find are sociological, focusing on people's perception on it and on agreement between experts rather than the actual subject - It's so I don't get intermediate
298.
by gobdovan 3y ago
Devices - thermionic, Cooling - electronic, Atomic vibrations? Number of applications! 3-D integration and space exploration
299.
by gobdovan 3y ago
True, and it was even more accurate a few years ago. Now Google moved towards understanding queries and giving context-specific answers rather than a strictly keyword based engine. You were considered a bad googler if you asked it a questio
300.
by gobdovan 3y ago
maybe streamsync [0] comes close to what you describe? "Streamsync is an open-source framework for creating data apps. Build user interfaces using a visual editor; write the backend code in Python." [0] https://github.c

More ›