theptip

17 ms·

391.
by theptip 11mo ago
> just prompt engineering This dismisses a lot of actual hard work. The scaffolding required to get SOTA performance is non-trivial! Eg how do you build representative evals and measure forward progress? Also, tool calling, caching, etc
392.
The Agent Lab Thesis(latent.space)
1 points by theptip 11mo ago|0 comments
393.
by theptip 11mo ago
What’s SOTA in this area right now?
394.
by theptip 11mo ago
Sure, ideally we can decouple the provider implementation and use a yubikey-type device if we want, or let the OS Secure Enclave handle it for the 99% of users that don’t care. The main point is it should be a protocol from the PoV of the c
395.
by theptip 11mo ago
> users would be able to control others from central browser controls that apply to websites broadly. Great to see this finally. It’s obviously the way it should have been implemented from the beginning. We still see this technically myo
396.
by theptip 11mo ago
> So did they start from scratch with this one Their major version number bumps are a new pre-trained model. Minor bumps are changes/improvements to post-training on the same foundation.
397.
The AI water issue is fake(andymasley.substack.com)
34 points by theptip 11mo ago|39 comments
398.
by theptip 11mo ago
I think this is kind of like saying “Only three kinds of internet products work, SaaS, webpages, and mobile apps” At the level of granularity selected, maybe true. But too coarse to make any interesting distinctions or predictions.
399.
by theptip 1y ago
Seems like you’ve smuggled in your conclusion with your definition of “highest value”? Using the normal definition (how many dollars we will pay for it) the statement is of course objectively false. Agriculture is certainly necessary, but t
400.
by theptip 1y ago
Intelligence is the manifold that these brute-force algorithms learn. Of course we don’t brute-force this in our lifetime. Evolution encoded the coarse structure of the manifold over billions of years. And then encoded a hyper-compressed me
401.
by theptip 1y ago
Elon did some napkin math along the lines of “we will make tunnel boring ~10x faster by: 4x lower boring surface area, 4x faster boring because the industry are currently idiots not operating their machines at the limits that physics dictat
402.
by theptip 1y ago
It seems pretty obvious that the bar needs to be raised. > A security report lands in your inbox. It claims there's a buffer overflow in a specific function. The report is well-formatted, includes CVE-style nomenclature, and uses ap
403.
by theptip 1y ago
You can also cache the DB across CI runs, which will reduce the randomness (ie failures won’t just disappear between runs).
404.
by theptip 1y ago
Also not an expert :) I thought plasticity is an O(hours-days) learning mechanism. But I did some research and there is also Short Term Plasticity O(second) [1] which is a crucial part of working memory. We'd need that. But it seems it
405.
by theptip 1y ago
Do you have any particular brain systems in mind that are essential for consciousness and also require mutable state?
406.
by theptip 1y ago
Self-play for memory/skill consolidation, plus flushing out metabolites which can’t be cleaned while consciousness is online?
407.
by theptip 1y ago
> a high chance that our brains require "write access" to function There are multiple learning mechanisms that happen on different time-frames, eg neural plasticity, hippocampus are both longer-term processes for memory consoli
408.
by theptip 1y ago
> But; they don't learn If we took your brain and perfectly digitized it on read-only hardware, would you expect to still “think”? Do amnesiacs who are incapable of laying down long-term memories not think? I personally believe that
409.
by theptip 1y ago
Bug hunting tends to be interpolation, which LLMs are really good at. Writing code is often some extrapolation (or interpolating at a much more abstract level).
410.
by theptip 1y ago
This exact problem comes up in AI alignment. It’s not enough to just look at the legible outputs. If you are going to trust someone with important responsibilities, you want them to “show their working” and convince you that that are not fa
411.
by theptip 1y ago
I think the linked https://blog.gregbrockman.com/figuring-out-the-cto-role-at-s... is much more interesting, it gives more actionable detail and advice. As Brockman says, you need a very strong VP Eng to make this possible.
412.
by theptip 1y ago
So, what’s the state of the art in guided state-space exploration/fuzzing? Seems if you have a reference implementation your fuzzer should be able to do some nice white-box validation to ensure you are behaving the same as the old impl
413.
by theptip 1y ago
It’s not a good safety valve though - sure, in a 2-party system, one of the parties can be taken over from within. But it’s uncommon and hard to achieve, and risks alienating voters while the civil war is going on. On the other hand, with I
414.
by theptip 1y ago
Now we are talking. And the dynamic that makes political parties so toxic IMO is “first past the post” voting. If it’s your team or the “worse” team, you tolerate any flaw in your team. If there was a pressure valve where another party can
415.
by theptip 1y ago
> Because our enhanced Session class now enables retries by default, we don't even need to instantiate it directly. modshim's AST rewriting ensures that internal references within the requests module are updated. This means con
416.
by theptip 1y ago
My take on AI-for docs is - it’s good. But you need to have a human review. It’s a lot easier to have someone who knows the code well review a paragraph of text than to ask them to write that paragraph. Good comments make the code much easi
417.
by theptip 1y ago
We can’t know, but that scenario would support my point.
418.
by theptip 1y ago
Humanoid robotics (ie replacing the majority of workers) is highly software-limited right now. Here’s a napkin-sketch proof: for many decades we have had hardware that is capable of dextrously automating specific tasks (eg car manufacture)
419.
by theptip 1y ago
Why does it have to be a dichotomy? Raise money from AGI-pilled investors with an AGI pitch. Raise more money from AGI-skeptics with a B2C growth story. Moar money == moar GPUs.
420.
by theptip 1y ago
This premise is a bit silly. If the machine god gets bored it can just create new life in its own image.

More ›