This sort of thing is really useful for helping people to understand what the purpose of scientific publication is. It's not about presenting finished products to society, it's more like a Discord where you communicate new findings with other authors. Often the reason for a communication is because you found something weird, and you want other people to know about it so they can help you confirm or rule it out as bad data. People shouldn't feel gunshy about this. (My field even has a conference for failed results, CFAIL.) I like to highlight examples of this stuff, because I see so many angry online comments when a paper turns out to be "wrong" or doesn't replicate.
DraftKings alone is estimated to do $6.5bn in revenue in 2026, and the whole US segment is close to $20bn. This is the market that Kalshi et al. are targeting, and these are huge sums of money. You can make sports betting artificially look small by comparing it to the entire equities market (but why?!), however the more reasonable standard of retail product/entertainment companies these are big and rapidly growing businesses.
Sports betting was, if not prohibited, then highly-proscribed for most of my life. You could do it in a few places like OTB, go to Vegas, or deal with an illegal bookie. People still gambled, but this made the problem orders of magnitude smaller than it is today. We did not need Sharia law or the guillotine, just an existing system of laws that were generally highly workable. It is not helpful to pretend like we need wild solutions just to achieve results we’ve already achieved and maintained for decades.
The hope of SMR is that it can be cost-effective at the same level that other factory-produced power sources are. If that's not on the table, then large-scale nuclear deployment mostly isn't going to happen.
Maybe, but I think this misstates how security vulnerabilities work. Vulnerabilities are logic flaws, and we have every reason to believe that there is a maximum number of such flaws in any given system that allow exploitation; and even that we can conceivably develop logic that excludes any flaws. Whereas weapons and armor are devices that deliver and deflect/absorb energy, and any increase in the power of one means we need a corresponding increase in the other.
Now maybe our understanding of logic systems is wrong, and it's just fundamentally impossible to develop programs that lack exploitable vulnerabilities -- that you can always "exploit with more energy". But there's no reason to believe the energy metaphor transfers to logic and intelligence.
Once you have a sustainable process, these last holdout areas like dredging (where you can't use batteries or wires) can have relatively terrible EROI as long as they're a small minority of the economy. All aviation and shipping combined is about 6% of final energy, and with major redesigns for shipping it could be as low as 4%. So you'd trade a much better EROI for 96% of the economy for a terrible one that only affects 4%.
This is already how Medicare works and it covers 70 million Americans. I’m not aware of any of these hypothetical violent outcomes being a huge problem for the program.
I was an introvert to the point where I started getting my hair cut at a place that plays movies, just to avoid awkward conversation with my hairdresser. Post-pandemic, I love talking to barbers and cab drivers, even if they have crazy ideas about the world. I wasn’t even socially starved, but just that limited period of socialization made me hungry for human interaction.
The Federal laws on grants are Byzantine and I’m not going to defend them. I think they make research much too expensive, and the compliance costs exclude smaller businesses and institutions. With that said: the general public voted for them. They did this because politicians raised the specter of fraud (which is real) and the public does not like their tax dollars being stolen. The current administration is doing things right now that sure look a lot like the type of stuff that motivates even stricter laws in the future, and will cause blowback for years. So no, I don’t think they’re in a good position to leave us with “reforms”. If anything I think they’re going to be gone in two years and all their works will sum up to even more new laws designed to keep the next version of them in check.
I’m genuinely not an expert, but isn’t the beauty of MoE models the fact that we explicitly don’t evaluate every parameter on inference? We evaluate exactly the subset that are needed to evaluate a prompt. Seems like this will bring back data-dependent branches again.
Some of this overhead is nonsense. A lot of it is training and fellowships. Some of it exists because the government has strong regulations to prevent self-dealing and fraud. I certainly think some of those regulations could be relaxed, but I’m not sure this administration would produce better ones.
What I think should happen is that these schools should increase their class sizes to match demand, rather than keeping them artificially low in order to game acceptance criteria. I think it’s a systematic societal failure that we’ve created the incentives for educational institutions to optimize for these metrics and exclusivity, rather than to educate people.
Author of the blog post here: just for the record, I did try replaying against other models. I replayed from GPT 5.5 to 5.5-mini. It "worked" (replay not rejected) but mini didn't hand me the plaintext of the reasoning. I'm not a jailbreaking expert, so I didn't really "push" on it very hard. I also didn't pursue Opus/Haiku for reasons I don't really recall. These folks took it all the way. I'm bummed it didn't work for me, but having these folks prove the channel out is a very good consolation prize.
They started stockpiling in 2024. The estimate is that they had 1.4 billion barrels by the time the Iran war started, and they're down to 1.2 billion now. (Nobody knows for sure, though.) Meanwhile the US has 304 million barrels left in our reserves, but there's a (disputed) operational minimum around 225 million, where draw slows down and the caverns start to suffer damage. (That's about 90 days at current draw rates. Also, this is a loan program, where oil needs to be "repaid" with interest starting in November, so that's going to affect prices too.)
Oil producer != oil power. What matters is the ability for a government to routinely influence prices. The US government could, in theory, ban all exports or manually dictate production but we tend not to do that (yet). Our main government price-control lever comes from SPR releases, but we're nearly tapped out. China has no real legal restrictions on what its government can dictate except for what makes sense geopolitically, so at any moment it can make oil vastly cheaper or more expensive, even moreso than OPEC can by manipulating supply.
I own a 2018 Model 3 and I love it. Plus it's in good shape with limited maintenance. I'm looking to trade it in. But honestly, I don't want to buy effectively the same car but slightly worse, with no turn signals (the recent M3) or an uglier version (the Model Y.) Historically, upgrading from an eight-year old car to the current version of the same model was a pretty big upgrade. Not with Tesla: I'll either keep this car until the wheels fall off, or I'll buy an R2.
"we will provide up to two weeks’ worth of tokens to facilitate your continued operation during the job search. We also partnered with ThriveFlow to furnish, as an option, a collection of expertly-crafted grief counseling prompts" seems like it.
I don't know what a vibe coder is to you, but I expect there are dozens of experienced software developers on HN who are willing to throw a month of work over the course of a year (and a few hundred dollars of Codex and Claude Code subscriptions/tokens they're already paying for) to write customized software for their friends and loved ones to use. And that while many of these tools will be mediocre or won't get maintained, a non-trivial number will be open sourced and will form a new layer of competition with existing SaaS businesses. (And even if I'm wrong, we're barely a year into the era of coding agents. Wait til 2027 to make big predictions.)
I’ve spent the last three days writing an inventory management system for a small store, using Codex and Claude Code. I realize that fully replacing online tools like Shopify is a lot more work than that (the payment integrations particularly), but: I sure wouldn’t want to be in the SaaS business right now.
They can't do this at the same margins. Cloud compute traditionally has <= 30% margins, and that's the very generous profit margin that folks like AWS are able to squeeze out of it with lock-in and proprietary tooling (something that will probably go away as software becomes more commoditized.) Frontier model companies (e.g., SpaceX) have been promising profit margins in the 75%+ range.
Subagents are a godsend. Frontier models for initial planning and spec writing, cheaper models for implementing from the spec, frontier models for review. Frontier models to help with research. The availability of cheaper models is great, since it gives more price points for different tasks.
A lot of this is downstream of compensation schemes that explicitly reward dumb metrics, like raw paper-count without subjective evaluation of contribution or quality. I don't want to generalize, but this seems to be more common in countries that are not the US or Europe.