Is it "sensible oversight" to shoot first and ask questions later—to shut down all payments on a whim and then implement a review process (assuming it actually exists) afterwards which will be immediately overloaded?
You might as well starve prisoners and then say sensible calorie management is not in and of itself foolish.
There's a methodological flaw here—you asked the model to generate an example, and then simulate the results on that example. For all you know, that specific (very simple) example and its output are available on the internet somewhere. Try it with five strings of other random words and see how it does.
I think there's a number of issues with this diagnosis, but chiefly: Trump won a second term on the back of a massive surge in non-white support. He's basically where he was in 2016 with white voters.
That's a morally questionable action—it might encourage some businesses to stay on Windows 98 even longer. And if that weren't bad enough already, the employees would also have to use Teams!
But what’s missing is just a fundamental sense of scale. One can rightly think this is a bit ridiculous, while also understanding it’s not that big of a deal. Honestly, there are so many economic and social issues that have real importance on people’s lives, and half the country is wound up in a culture war over which words are considered polite or not.
The title is wrong; prevention is what they do. The actual quote from the article is:
> As a result, we prevented 2.36 million policy-violating apps from being published on Google Play and banned more than 158,000 bad developer accounts that attempted to publish harmful apps.
I have done the latter much more than the former. My experience has been the issues come from inputs that you don’t foresee, not reliability on in-distribution uses (which would be your “training” data for prompt optimization). And the worry is that this kind of optimization would lead to substantive revisions of the guidelines set out in the prompt, which could further compromise performance out of distribution.
To the extent that you need to eke out reliability on the margins, one is vastly better served by actual fine-tuning, which is available both for open-source models and most major proprietary models.
I always find myself baffled by “prompt optimization” frameworks. Do people really find themselves needing random perturbations of a fixed prompt to improve accuracy? It’s my experience that the challenging part of writing a prompt is figuring out what the task you want done is, and understanding which data you need to pass to the model to make the task achievable. None of that can be achieved by “optimizing” the prompt—the hard part is a layer of abstraction upward.
Yeah, I think NVDA could fall another 20% and its stock price would still imply "massively profitable company primed for continued explosive growth." Its previous valuation was a premium even above that story.
If I’m reading this correctly, it isn’t a severance package, exactly. The only “compensation” an employee is getting in exchange for agreeing to resign in September is the ability to continue working from home. Things are messier if agencies start placing those employees on paid administrative leave etc.
Not really - there's a difference between having the docstring of a function available for you to read, and a model which has learned from thousands of examples how to use a particular API and integrate it into a larger set of instructions. The latter is vastly faster and takes much less human work than the former.
The New York Times is relying on their audience's ability to understand that plants are not conscious beings. "Mournful cries" is just an evocative way of describing the auditory signals plants produce under stress—which the article makes clear. I can't understand what's so "deranged and insane" about some stylistic flair.
I don’t agree with that analysis. The modern Supreme Court loves to wade into high-profile First Amendment issues, all the more when it’s actually a relatively novel application. They’ll take this case either to be argued at the end of this term or next term, and it will be subject to their analysis and whatever Trump decides to do with his discretion.
I’d note that most U.S. law journals work very differently from what you describe. Strict Bluebook compliance is their raison d'être, much more than legal scholarship. I am not exaggerating that. (And authors are not generally expected to get it right themselves, fleets of law student editors do the checking and correcting.)
Later on, it also says that Apple was stopped work on the second-generation Vision Pro and is focusing only on a lower-cost device. So that’s actually pretty reasonable.
I don’t want to be in the business of LLM defender, but it’s just hard to imagine this aging well when you step back and look at the pace of advancement here. In the realm of “real math and science”, O1 has improved from 0% to 50% on AIME today. A year ago, LLMs could only write little functions, not much better than searching StackOverflow. Today, they can write thousands of lines of code that work together with minimal supervision.
I’m sure this tech continues to have many limitations, but every piece of trajectory evidence we have points in the same direction. I just think you should be prepared for the ratio of “real” work vs. LLM-capable work to become increasingly small.
I think what I’m saying is a bit more nuanced than that. LLMs currently struggle with very “wide”, long-run reasoning tasks (e.g., the evolution over time of a million-line codebase). That isn’t because they are secretly stupid and their capabilities are all hype, it’s just that this technology currently has a different balance of strengths and weaknesses than human intelligence, which tends to more smoothly extrapolate to longer-horizon tasks.
We are seeing steady improvement on long-run tasks (SWE-Bench being one example) and much more improvement on shorter, more well-defined tasks. The latter capabilities aren’t “hype” or just for show, there really is productive work like that to be done in the world! It’s just not everything, yet.
It’s undeniably less impressive than a human on the same task, but who cares at the end of the day? It can do 10,000 attempts in the time a person can do 1. Obviously improving that ratio will help for any number of reasons, but if you have a computer that can do a task in 5 minutes that will take a human 3 hours, it doesn’t necessarily matter very much how you got there.
It’s definitely the case that there are some programming workflows where LLMs aren’t useful. But I can say with certainty that there are many where they have become incredibly useful recently. The difference between even GPT-4 last year and C3.5/GPT-4o this year is profound.
I recently wrote a complex web frontend for a tool I’ve been building with Cursor/Claude and I wrote maybe 10% of the code; the rest with broad instructions. Had I done it all myself (or even with GitHub Copilot only) it would have taken 5 times longer. You can say this isn’t the most complex task on the planet, but it’s real work, and it matters a lot! So for increasingly many, regardless of your personal experience, these things have gone far beyond “useful toy”.
A lot of skepticism here, but these are astonishing results! People should realize we’re reaching the point where LLMs are surpassing humans in any task limited in scope enough to be a “benchmark”. And as anyone who’s spent time using Claude 3.5 Sonnet / GPT-4o can attest, these things really are useful and smart! (And, if these results hold up, O1 is much, much smarter.) This is a nerve-wracking time to be a knowledge worker for sure.
This is the purpose of Zelle, Venmo, money wires, and checks. But there are many problems they don’t solve, that customers and sellers prefer to be solved and are willing to pay for.
GitHub Copilot's telemetry does collect data on whether generated code snippets end up staying in the code, so presumably models are tuned on this feedback. But you haven't solved any of the problems set out by Karpathy here—this is just bankshot RLHF.
If you’re concerned about the environment, that is a trade you should take every time. AI is 100-1000x more carbon-efficient at writing (prose or code) than a human doing the same task. https://www.nature.com/articles/s41598-024-54271-x
I agree with most of the criticisms here, and will add on one more: while it is generally true that you can’t beat “serverless” inference pricing for LLMs, production deployments often depend on fine-tuned models, for which these providers typically charge much more to host. That’s where the cost (and security, etc.) advantage for running on dedicated hardware comes in.