theptip

32 ms·

661.
by theptip 2y ago
> However, before making policy decisions, I would want the second order data: I think this the wrong lens. The correct lens is: if they don't voluntarily make this change, will they be forced to? The incoming administration seems c
662.
by theptip 2y ago
I think what they are saying there is the press release interpretation of experiments showing a false positive rate of 10-20%, with error bars wide enough that stating a percentage gives too many significant figures. But the definition of F
663.
by theptip 2y ago
Community notes seems to be quite well received. I like that the algorithm seems to be public and (IIUC) tamper-evident. The obvious context is that either Meta gets out of the content moderation game voluntarily, or the incoming admin goes
664.
by theptip 2y ago
How much would HTMX internals change if these proposals were accepted? Is this a big simplification or a small amount of what HTMX covers? Similarly, any interesting ways you could see other libraries adopting these new options?
665.
by theptip 2y ago
Isn’t the cloud GPU market covering this? I can run a model for $2/hr, or get a 8xH100 if I need to play with something bigger.
666.
by theptip 2y ago
$3k for a 128GB standalone is quite favorable pricing considering the next best option at home is going to be a 32GB 5090 at $2k for the card alone, so probably $3k when you’re done building a rig around it.
667.
by theptip 2y ago
That's... not an ad? User used Meta AI to touch up a selfie, now Meta AI is generating more selfies for you. I can understand the annoyance from cross-linking apps in the same way it's annoying to get the Threads popup in your fee
668.
by theptip 2y ago
Some good observations here but I disagree with the conclusion. > Shipping features from the roadmap is a chunk of our core job. It shouldn’t usually be in our OKRs. First of all, your roadmap shouldn’t be static. You should give teams f
669.
by theptip 2y ago
Karpathy discussed this at length on the Lex Fridman podcast: https://lexfridman.com/andrej-karpathy/ IIRC I think it’s the section (1:23:25) – Camera vision The TL;DR is that sensor fusion is really hard, and their be
670.
by theptip 2y ago
Right. As well as, or in able to. Not instead of.
671.
by theptip 2y ago
What if socializing IRL is something you can practice with pro-social bots? Or, what if depressed people find the pressure of consequences too stressful to “just go outside and make friends”, and so the realistic options are chatbot or not
672.
by theptip 2y ago
No, it’s reasonable. If your team uses Git then it’s a valid question to establish if someone has only worked with Perforce. They didn’t say how heavily they weight the question. (All that said I expect that, soon, experience with the appro
673.
by theptip 2y ago
The context here is super-important - the commenter is the author of Redis. So, a super-experienced and productive low-level programmer. It’s not surprising that Staff-plus experts find LLMs much less useful. Though I’d be interested if thi
674.
by theptip 2y ago
A mixer can run at the full speed of its input belts (ie given two full belts input, it will output two full belts). In the naive solution you are effectively allocating a single belt output from each layer-1 mixer, so the system will run a
675.
by theptip 2y ago
Folks like to claim that software is Engineering but it’s as much Craftsmanship. Hence, taste is in fact important. In some areas you need more engineering but API design, for example, is mostly taste and hardly any science.
676.
by theptip 2y ago
Great breakdown of the "architectural decision log" for the evolution of this system. > This model broke down when we added backpressure and resilience patterns to our agents. We faced new challenges: what happens when the thir
677.
by theptip 2y ago
HTML version of the paper itself: https://arxiv.org/html/2412.14093v2
678.
AIs Will Increasingly Fake Alignment(thezvi.substack.com)
104 points by theptip 2y ago|106 comments
679.
by theptip 2y ago
The more/smaller microservices you have, the more frequently your APIs must change. It’s more fruitful to recognize that this is a dimension of freedom rather than a binary decision.
680.
by theptip 2y ago
The comparison is with a monolith where versioning isn’t required at all, since you can refactor all clients atomically in one change.
681.
by theptip 2y ago
This is baffling. I’m sure there are many technical reasons I don’t grok that AMD’s job is challenging, but it’s wild that they are dropping the ball on such obvious stuff as this. The prize is trillions of dollars, and they can print hundr
682.
by theptip 2y ago
Not a silver bullet; you increase api versioning overhead between services for example.
683.
by theptip 2y ago
Agreed that the flag seems highly dubious here.
684.
by theptip 2y ago
Well, now you’ve moved the goalposts from “learn anything” to “learn at human level”. Sure, they don’t have that yet.
685.
by theptip 2y ago
> LLMs aren't really capable of "learning" anything outside their training data. ChatGPT has had for some time the feature of storing memories about its conversations with users. And you can use function calling to make th
686.
by theptip 2y ago
Thanks for clarifying. I do still think the paper shows what should be considered “intentionality”. The Chain of Thought clearly shows explicit planning, considering whether to subvert safety systems, and considering how to avoid being caug
687.
by theptip 2y ago
The prediction from the LessWrong folks was that it’s inevitable that a rational actor with “goals” would do this, so models trained purely on RL would exhibit this too. (Instrumental Convergence is the name for the theory that predicts pow
688.
by theptip 2y ago
Intention is irrelevant, what matters is the observed actions. In this case you can see the chain of thought and the model is clearly ruminating on whether to subvert the RLHF process. This has nothing to do with left-polarization / wo
689.
by theptip 2y ago
But there is substantial guidance. RLHF, constitutional AI, etc. Every lab is trying to steer their models to behave in pro-social ways, and nobody has found a way to make this guidance stick reliably.
690.
by theptip 2y ago
Simply, as LLM capabilities increase, writing code might disappear, but building products will remain for much longer. If you enjoy writing code, you might have to make it a hobby like gardening, instead of earning money from it. But the br

More ›