lukev

28 ms·

391.
by lukev 2y ago
Lots of comments in this thread are reacting against the idea that writing is required for clear thinking. I think this is true, but a little more nuanced. Writing and clear thinking are are related in at least two ways: 1. Both good writin
392.
by lukev 2y ago
Multiple things can be true at the same time. 1. Many companies have too many layers of management, with too much abstraction from the actual product, which kills drive and focus. 2. Micromanagement and disempowerment of employees is bad ma
393.
by lukev 2y ago
I dunno, I'm 6'2" with corresponding hand size and I'm in the "won't go larger than a 13 mini" camp. I think preference probably plays a bigger role than size. I see a lot of tiny people manhandling pros a
394.
by lukev 2y ago
I'm not speculating here, I am telling you how it works. Phones ARE banned from school and the kids DO socialize in person there. But then they all get home and stay in touch. Some have phones, some have tablets, some just their school
395.
by lukev 2y ago
It's very easy to say this if you're not a parent. Social media does some harm. Being socially excluded does far _more_ harm. Which starts to happen in middle school. So as a parent, you have three choices: 1. Deal with the emotio
396.
by lukev 2y ago
The cool thing about the various verions of UUID is that they're all compatible. The differences almost all come down to database locality (and therefore performance.) The exception is if you're extracting the time portion of a ti
397.
by lukev 2y ago
I don't even think that's what this is doing, though. As technologists, the point isn't to be bullish or bearish on LLMs... we should be focused on empirically understanding what they can do, and why, so that we can design sy
398.
by lukev 2y ago
Interesting. I agree with you about being most comfortable with a desktop/keyboard interface, but have the exact opposite opinion regarding macs. IMO, OSX is the perfect platform for a keyboard-driven power user. It's unix/BS
399.
by lukev 2y ago
Also, making neural networks faster/cheaper is a big part of how they advance. We've known about neural architectures since the 70s, but we couldn't build them big enough to be actually useful until the advent of the GPU. Sim
400.
by lukev 2y ago
I never said it was. FYI, whataboutism in a thread about Nazis makes it look a little like you're defending Nazism. Was that your intent? If not, what did you intend to add to the conversation with this (true) observation?
401.
by lukev 2y ago
Fascists don't believe that, and will happily kill people with different ideas if they get the power to do so. This is the paradox of tolerance. It's pretty well understood, except apparently on online message boards where there a
402.
by lukev 2y ago
No matter which version of the story is actually true, they are all tragic. One important takeaway is mentioned in the article: > None of this excuses the treatment of Turing during his final years, says Prof Copeland. The injustice was
403.
by lukev 2y ago
The premise of a startup is that your investors expect you to mature and build something valuable. If you want to be a "funny little guy," stick to hackathons.
404.
by lukev 2y ago
This is only true if you are very uncreative about names.
405.
by lukev 2y ago
Was doing a bit of searching to get context on this story and found an unrelated but interesting tangent... there's another unrelated startup named pear.ai that's been around for years. It's in a different domain, but the nam
406.
by lukev 2y ago
The problem is when it’s an app that sometimes you want notifications from. I deleted Uber, for example, when it started spamming me.
407.
by lukev 2y ago
Oh they are 100% losing money hand over fist if you include training costs and the eye-watering salaries they pay some of their employees. I was responding to someone upthread suggesting that they were running even inference at a loss.
408.
by lukev 2y ago
All these efficiencies just increase OpenAI's margin on inference. Of course it's not "one cluster per customer" and of course a customer can't saturate a cluster by themselves, my illustration was only to point out
409.
by lukev 2y ago
My point is that at current API pricing the users are paying enough to cover inference costs.
410.
by lukev 2y ago
Obviously, that's my point. We can do the math. GPT-4o can emit about 70 tokens a second. API pricing is $10/million for output tokens and $2.5/million for input tokens. Assuming a workload where inputs tokens are 10:1 with o
411.
by lukev 2y ago
Model training is what costs so much. I would expect OpenAI makes a profit on inference services.
412.
by lukev 2y ago
If you care about this, let's take cryptocurrency offline first. Crypto still uses many multiples of the power consumed by AI, and unlike AI it can _never_ improve and provide more utility per watt since it is inefficient by design.
413.
by lukev 2y ago
Yes, the "surplus compute power" story doesn't hold water. The magic of AWS is in its software and control plane.
414.
by lukev 2y ago
Not at all! I am broadly speaking very pro-tech. What I am against is the “if it’s tech it’s good” mindset that seems to have infected far too many. I mention crypto because it’s the largest example of tech that is not good for the world.
415.
by lukev 2y ago
The utility of LLMs clearly exists (I'm building a product on this premise, so I'm not uninterested!) But hype also exists. How closely they are matched is not yet clear. But your comment seems to indicate that the "pro-tech&
416.
by lukev 2y ago
Yes. And another variable is what humans believe state of the art AI to be capable of. If you dropped ChatGPT in 2015 I’m sure it would pass the Turing test quite a lot. Much less so now that people are more familiar.
417.
by lukev 2y ago
Yes indeed, and personally if we have AGI I believe it will arise from multiple LLMs working in tandem with other types of machine learning, databases for "memory", more traditional compute functions, and a connectivity layer betw
418.
by lukev 2y ago
Ok but this presses on a latent question: what do we mean by alignment? Practically it's come to mean just sanitization... "don't say something nasty or embarrassing to users." But that doesn't apply here, the reaso
419.
by lukev 2y ago
It's interesting to note that there's really two things going on here: 1. A LLM (probably a finetuned GPT-4o) trained specifically to read and emit good chain-of-thought prompts. 2. Runtime code that iteratively re-prompts the mod
420.
by lukev 2y ago
Look, I've been experimenting with this for the past year, and this is definitely the happy path. The unhappy path, which I've also experienced, is that the model outputs something plausible but false but that aligns with an area

More ›