Great catch. Super disappointing that AI companies continue to do things like this. It’s a great result either way but predictably the excitement is focused on the jump from o1, which is now in question.
To me it's very frustrating because such little caveats make benchmarks less reliable. Implicitly, benchmarks are no different from tests in that someone/something who scores high on a benchmark/test should be able to generalize that knowledge out into the real world.
While that is true with humans taking tests, it's not really true with AIs evaluating on benchmarks.
SWE-bench is a great example. Claude Sonnet can get something like a 50% on verified, whereas I think I might be able to score a 20-25%? So, Claude is a better programmer than me.
Except that isn't really true. Claude can still make a lot of clumsy mistakes. I wouldn't even say these are junior engineer mistakes. I've used it for creative programming tasks and have found one example where it tried to use a library written for d3js for a p5js programming example. The confusion is kind of understandable, but it's also a really dumb mistake.
Some very simple explanations, the models were probably overfitted to a degree on Python given its popularity in AI/ML work, and SWE-bench is all Python. Also, the underlying Github issues are quite old, so they probably contaminated the training data and the models have simply memorized the answers.
Or maybe benchmarks are just bad at measuring intelligence in general.
Regardless, every time a model beats a benchmark I'm annoyed by the fact that I have no clue whatsoever how much this actually translates into real world performance. Did OpenAI/Anthropic/Google actually create something that will automate wide swathes of the software engineering profession? Or did they create the world's most knowledgeable junior engineer?
> Some very simple explanations, the models were probably overfitted to a degree on Python given its popularity in AI/ML work, and SWE-bench is all Python. Also, the underlying Github issues are quite old, so they probably contaminated the training data and the models have simply memorized the answers.
My understanding is that it works by checking if the proposed solution passes test-cases included in the original (human) PR. This seems to present some problems too, because there are surely ways to write code that passes the tests but would fail human review for one reason or another. It would be interesting to not only see the pass rate but also the rate at which the proposed solutions are preferred to the original ones (preferably evaluated by a human but even an LLM comparing the two solutions would be interesting).
If I recall correctly the authors of the benchmark did mention on Twitter that for certain issues models will submit an answer that technically passes the test but is kind of questionable, so yeah, good point.
I felt this same way as image generation was rapidly improving, but I've been caught by surprise and impressed with how resilient we have been in the face of it.
Turns out it's surprisingly, at least for me, to tune out the slop. Some platforms will fall victim to it (Google image search, for one), but new platforms will spring up to take their place.
What frustrates me the most about OpenAI is that as recently as this summer they were talking non-stop about how gigantic $100 billion models are all you need for AGI and that it’s just a matter of time until we reach this scale. And if you didn’t see this you’re a simpleton who doesn’t understand how exponential curves work.
And then all of the sudden o1 comes out and the narrative from them has shifted entirely. “Obviously massive models aren’t enough to get to AGI, but all we have to do is scale up inference time compute and we’ll get there!” And if you don’t see this you’re just a simpleton.
I wish that OpenAI was called out for this shift more often. Because I haven’t heard even one of their employees acknowledge this. At some point you have to just ignore them until they start actually publishing science that supports their beliefs, but that won’t happen because it doesn’t generate revenue.
> that won’t happen because it doesn’t generate revenue.
OpenAI made real progress towards a computational understanding of human language and cognition. I'm sorry they have become a for-profit entity (the paperwork lags behind reality, of course). A fiduciary duty does not serve humanity. The quality and credibility of their communications have fallen dramatically.
> how gigantic $100 billion models are all you need for AGI and that it’s just a matter of time until we reach this scale. And if you didn’t see this you’re a simpleton who doesn’t understand how exponential curves work.
OpenAI made no such claim. LLM stans on the internet definitely made such claim, that the Stargate project would be AGI and whatnot. But, like crypto bros, GPT hyperfans are just constantly deluded/lying so you shouldn't project their claims onto the corporations they simp for.
That being said, Anthropic's CEO made a claim closer to what you're saying, that a 10-100 billion dollars model would be better than a human in almost every way.
America gave a trillion dollars out so a lot of people can have 1500 dollars. We have enough money, and we are all not going to live forever. I don’t know what the hold up is.
It’s not an exact figure, but plenty of employees claimed that after we train a model with 2-3 orders of magnitude of compute over GPT-4 we’d reach AGI, which puts us at $10-$100 billion.
See Situational Awareness for one example of this.
This summer he made a comment saying "we can make GPTs 5,6,7 more powerful" and someone editorialized it as being "via parameter count/scale" which he didn't say.
The loudest opinions surface. I wouldn’t take lack of public sympathy to mean that the average American condones this type of behavior.
I am no more sympathetic than I am towards the many people killed each day in the US, but I think the assassination and the subsequent justification of it because he was a small cog in the system is abhorrent. And I hope that the majority of the country feels the same way.
Whether or not the assassination is justified, I don't think it makes sense to call him a small cog. He's the leader of one of the largest health insurance companies in America.
As much as I hate to point it out, it’s not so simple.
He had a legally enforceable mandate to maximize shareholder value. There’s some wiggle room in how that’s accomplished, but not as much as it seems from the outside.
That does make him, much as I hate to say it, a small cog.
It’s funny to me that so often on HN we have discussions about making ethical choices in who we work for and which projects we contribute to, yet the sentiment on this CEO is honestly the most sympathetic I have seen in any comment section
He chose to take the role. He profited from it massively and did more to perpetuate it than any other single individual at the company.
If we were talking about a developer who was writing software for drone bombs wiping out families, yeah maybe that’s a cog. But still there would be no shortage of judgement on that choice of profession
It’s plainly obvious that insurance companies are not good faith actors working to help people. They lobbied for this system, they are the beneficiaries of it, and they are reaping what they’ve sown
But no less deserving of being executed. His decisions led to the deaths of thousands. Justice — in the true sense of the word — demands an answer for this. If the systems we have in place won’t do it, then extrajudicial means are justified. And encouraging to see, frankly.
Just because his crimes were legal does not mean he should not face punishment.
> The loudest opinions surface. I wouldn’t take lack of public sympathy to mean that the average American condones this type of behavior.
I'm not sure I agree with you there. The indifference in some quarters, celebration in others, across either side of the political divide insofar as I've seen, leads me to believe that in some cases, the well of sympathy is running dry – or, worse, has run dry.
Elon and Vivek have no authority to close a federal agency, and Trump doesn’t either.
We’re currently in the bloviating stage of this election cycle; once attention dies down I fully expect DOGE to achieve very little and to die a slow death.
The sad part of all of this is that the government could absolutely be more efficient than it currently is, while still providing the same services. But that’d take serious thought and consensus building, which the incoming administration has no desire to engage in.
I’ll check it again in a year. Their “algorithm” is another polarization device, however. If you click or like anything, your feed fills with extremely similar things. Usually, that makes recovery from a closed mindset impossible.
Why does no one read the actual proposal before commenting?
It specifically states that this only applies to individuals with 80% of their wealth in tradeable assets. No founder is going to lose control because this doesn’t apply to them!
I can IPO, sell some shares and make billions but have to take 25% in tax or I can not do that and what happens? How do I turn my non IPO shares into profit? Presumably if you liquidate to cash you owe the capital gains right?
You’re still going to end up making a higher valuation on the stock market than you would trying other means to avoid this tax.
I'd love to. I'm genuinely interested in how this policy could be implemented and would love to read their suggestions. I think it's very hard to pull off successfully. Can you provide a link?
I thought Harris was adopting the President's 2025 budget proposal [1], which doesn't specifically state this is specific to tradable assets, but according to the downvoters I'm wrong about that. As far as I can tell it provides no comment on how "wealth" is determined.
I suppose the whole argument is moot anyway as the President doesn't pass a budget, Congress does. And this document is really about communicating priorities, not actual policy.
And if one wants to get really persnickety, Harris didn't actually say anything. Some people working for her campaign did.
Because that detail, like the $100milliom limit are irrelevant details subject to change. (Most) People have the ability to synthesize issue and are worried of the proposal's fundamental core:
"Do we want the government to tax unrealized gains?"
No. I find it very scary frankly, even though I believe that the top 0.01% of the US population are parasitical and their financial and political clout should be reined in.
Zoning and car-centric design are intimately connected, though. A walkable suburb just can’t support the number of cars that a typical western suburb supports today.
If you loosen zoning, what you’re going to end up with is denser development and less room for cars.
Streetcar suburbs are an example of this. You have more room than living in the core of the city, but you don’t have enough room for 3 F150s for the family.
The streetcar suburbs in my area have larger lots than the new suburbs they are building on the edge of town. Back in those days e*erpone had a garden and so lots were bigger.
of coures how the lot is built is different but it isn't size.
reply