jefftk

Where is 10% coming from? If you're focused on something far away shouldn't the fraction be <<1%?

I think you way over estimate how hard the knowledge bit is in a biological weapon ... if you just want to kill indiscriminately

That's a surprising claim: usually when I make this argument skeptics say that the knowledge barrier is so high that an LLM won't help enough!

I'd also be curious to hear what you think of the Aum Shinrikyo case, since that seems to me to be straightforwardly a failure of knowledge.

The idea that all that's stopping some disaffected Joe sitting on his sofa at home and suddenly deciding to make a biological weapon is he can't get a set of instructions from ChatGPT is to miss the point.

I agree motivation is a huge barrier, in the sense that almost no one would do it. But to keep catastrophic bio attacks from happening as the knowledge barrier decreases it's not enough that most people wouldn't have the motivation. If 0.0001% of people have the motivation and 0.0001% have the means then we're probably ok, but if 0.0001% of people have the motivation and 0.1% have the means then we're not (0.0001% * 0.1% * population > 1).

Don't buy the idea that the only thing that's standing in the way of people making bioweapons is the censoring of LLMs.

It's not a matter of a one thing standing in the way of people making bioweapons: there's a long chain of actions one would need to complete, and many places for the chain to fail. Access to expertise can reduce the chance of failure at many of these steps, and AIs can increasingly substitute for human expertise: https://securebio.org/benchmarks/

If you are motivated enough to assemble all the kit you'd need, and actually do it, then you should be motivated enough to find the knowledge to do it without chatgpt etc.

If someone was motivated enough to assemble the kit you need for anthrax and actually distribute it then you might expect them to also be motivated enough to find the knowledge to identify an appropriate strain, but in fact this is where Aum Shinrikyo failed in the 1990s: https://en.wikipedia.org/wiki/Aum_Shinrikyo#Incidents_before...

Lack of knowledge is one of many factors that can lead to failure, and LLMs make it less of a barrier.

biological weapons are a terrible tool to do what most people want to do - which is target specific enemies

Except:

1. There are also people out there who want to kill everyone. They're sufficiently rare that no one with the motive has also had the means, but as technological progress keeps lowering the bar the risk of motive and means intersecting increases.

2. This didn't stop the Soviets. They did an enormous amount of very dangerous research with minimal logical application.

Only infinitesimally more (most of what that chart is showing is that the growing degree days come earlier in the year), and it also gets colder and has a shorter growing season. I'd expect most plants to like this less.

I'm not sure the line defining what you would consider AI is actually all that bright. Sharpening? Content-aware fill? Seam carving? Style transfer? How would you define AI to draw the line where you want?

model capabilities across the board have only meaningfully improved in places where the labs are focusing their training efforts

That doesn't seem right. I use these models as research assistants when writing lots of random blog posts (including in economically ~useless areas like the history of contra dance) and Fable 5 is a serious improvement (when I don't get downgraded!) over Opus 4.6-4.8 which was a serious improvement over Opus 4.

release the models as open weights with a license (I.e no commercial use) due the high risk they present

For some kinds of risk (ex: walking people through on how to make infectious bioweapons) an open-weights approach would increase risk.

They're scraping to automatically update the wallpaper on their desktop. That's not something a website can do, even with fantastic UX.

Clocks used to be able to use the 60Hz cycle to track time, and grid providers would run slightly slow or fast ("time error correction") to get back into sync. A leap second would just be part of this.

I believe in the US this error correction has been discontinued in the East and in Texas, but is still done in the West for some kind of non-clock "inadvertent interchange" reasons I don't understand.

A Verizon tech called me yesterday and told me that they were not going to take Gizmohub down until they'd resolved these issues. They also walked me through getting set up, including getting me the 2FA code that was not coming through. Apparently delivery of 2FA codes to Google Fi has been an issue for them. After our call I saw that this 2FA code (but not the earlier codes) did actually come through under "Spam and Blocked", so if others are having issues I'd recommend looking there.

I know you put an asterisk, but to make it clear: these locations are very far from random. They're the most informative positions, which is why we check them.

"int radians" is a pretty strong code smell! I would be very worried about any struggle l autocomplete that followed something like that ;)

Free soda with large fry doesn't actually exist

I just searched and found free medium fries with a large soda at McD if you order through the app. Seems like bundling to me, and you have to use their particular app.

Anthropic being a big company means they could trivially hire an iOS team

Anthropic is primarily limited by how fast they can identify and onboard people who will be a good fit. Allocating that to ios would trade off against allocating that staffing growth elsewhere (ex: reliability, making the next generation model even better).

Why compare to what historical tooling (which is very different in terms of how it's built and what it does) instead of to value over replacement?

Substack handles emailing and payments well. Many people want to receive posts by email, which is a pain to do on your own with high deliverability. Some are willing to pay for it, and payments are similarly annoying (and benefit from a familiar interface, which Substack now provides).

Millions of people have signed up to receive blog posts by email with Substack. HCR alone has 3M subscribers. While AI replaces "here's common knowledge repackaged" blogs, people still care what the authors they trust think.

Now maybe sometime soon AI replaces that too, but I think by that point we're talking about "automation of the majority of human intellectual labor" and are well beyond blogs in particular.

No mention of Substack? Making money from paying subscribers has different trade-offs than making money from ads, but my read is that mostly traffic moved vs evaporated. But I do expect this to change further with AI, where as the author says, a blog needs to add something new and not just try to answer a question someone might search for.

There's also no discussion of how blogging has always been somewhat frothy: picking the successful blogs (by any metric) and then checking back later is almost guaranteed to show a decrease. A fair comparison would show the top blogs now vs then, or even better the overall landscape (but that's a ton of work).

With Google moving their smartwatch efforts to the pixel brand I'm worried about longevity here. It also has a lot of the downsides of the Gizmo: 20 contacts max, texting only through the app, can't interact with friends. If I'm going to get new ones for the kids I'll probably do a TickTalk or Cosmo that don't have these drawbacks.

There aren't enough phone numbers to make this work. There are only ~5.7B phone numbers in the NANP that can be assigned to individuals [1] and these countries have >400M people, so it's just ~14 phone numbers per person. Which works if we each have a couple, but not if we start handing out many per person.

[1] 10B ten-digit numbers, less prohibited ones like area codes starting with 0 or 1, the 555 exchange, etc.

Definitely! But the form factor is really important: since it's strapped to their wrist they don't lose it. Classmates with phone-shaped phones seem to loose theirs a lot.

Apple watch only works if a parent has an iPhone, and we're both on Android.

Not an option, unfortunately, since my wife and I are both on Android.

It was announced in April, but it was leaked in March (CMS bug) at which point external partners were already using it, and the most common rumored date for training competition is 2026-02-07 (I think Feb is likely, but that specific date is just rumor).

We lost ten days of the original plan (June 13th through 22nd) and got seven days (July 1st through 7th). You're not counting the portion of the original 14 days we already got.

Sacks has been out since March

I'm from Boston and I'd never heard of this, but reading a bit it sounds like Boston is unusual in not having pace runners.

Given how hilly the course is, this makes sense: running an even pace the whole distance is not a good plan.

Limiting exports of AI services to all foreigners is probably allowed under GATS, since there's no favoring one county over another. But even then there's a national security exemption, which fits reasonably well with US arguments here.

If a country thought the AI export restrictions were inconsistent with treaty the remedy is challenging them, not unilaterally imposing their own tariffs. But even if they got a favorable panel decision the US would speak, and the Appellate Body is non-functional because the US stopped consenting to the addition of new members and all the terms expired. Which means anything that gets appealed is frozen indefinitely waiting for the AB to reach a quorum that won't come until the US changes is mind (and Biden didn't reverse Trump's decision here).

You very likely know this, but to make it explicit: "US Persons" under ITAR is US Citizens + Lawful Permanent Residents (green card) + Protected Individuals (Non-citizen nationals like Samoans, Refugees/Asylumees). It doesn't include anything else, like H1B, TN, etc visas.

I would expect tickets issued during the first 40 days to be higher than later, as people haven't adjusted yet