timr

No. My argument is that many things in the world are hard in a way that computers cannot solve, and the existence of LLMs does not change that.

Synthetic biology is one of those things.

Not even close to the same thing. That's taking an existing viral transfection system and swapping in a single gene.

Thermo Fisher sells a kit to do this [1]. It's something you might do a few thousand times on the way to actually generating a novel virus.

> I’d argue estimate the risks of a synthetically induced pandemic of massive impact this century are very high

I'd argue that it probably already happened, and had nothing to do with AI.

[1] AAV transfection.

Yes, that's one of the ones I was talking about.

That's not a service that makes you a virus, and making one from an mRNA transcript is not trivial.

No, the sophomoric LLM argument does not apply here. The world is not a big computer.

The point is that the rate-limiting step in producing something infective is a couple million dollars in special lab equipment, skilled practitioners, and incredibly tricky and tedious procedures that take time and repetition to master.

So yeah, you can try to pay someone who has that expertise to make your top-sekret superbug (good luck with that), or you can do what any real threat actor is going to do, and put in the time and investment. That's the hard part, not asking ChatGPT for a genome sequence.

Let me put it another way: despite the synthesis pathways for many common illegal drugs being widely known and trivial enough that an undergrad O-chem major could do them, in fact doing this is a very good way to get caught and arrested. And that's easy.

No, you literally cannot. Synthesizing a stretch of DNA is trivial, and what you can send away for [1].

Building a functional virus is not even remotely a turn-key operation. Assuming you could get the genome, expressing that genome in a way that results in infectious virus is non-trivial.

[1] Even then, the companies will rarely give you long enough reads to make a full genome scaffold. It looks like there are two companies that will do 50-100kb assemblies as a service today, but I guarantee you they're going to look at what you're trying to make, and who is asking them to do it.

The design was never the hard part. You still have to make the thing.

Designing a modifcation to an existing virus is perhaps more achievable, but the implementation is still the hard part.

US in the 80s. I mean, I don't know if it was legal, but it happened.

Given that almost nothing underground in NYC is ever proactively removed when deprecated, I wonder how much of this system is left, and if any of it is actually visible today.

My mother was a pharmacist in a major city hospital, and I spent many days as a young child bumbling around the pharmacy while she worked. One of my memories of that time was being utterly fascinated by the tube system that they used to send IV bags and whatnot to the various floors.

Just fly everyone out in a group and have a big rock-paper-scissors tournament. You can save on airfare.

> I'm sorry, but if I'm doing the hiring, the person who isn't willing to come on site to do an interview will be one of the first people off the list to hire

Nobody said anything about not coming on site. That's how we used to do it, in the ancient pre-history of 2019. What I said was that if you bring someone on site only to leetcode them, you deserve not to hire anyone.

The downside, of course, is that you don't have the cheap, dumb pre-filter of a memorization test anymore, so you'll have to figure out a better way to winnow down the applicant pool whom you're willing to pay to bring on site.

Maybe finally software engineering will achieve a level of interview maturity seen in other professional fields...but I doubt it.

Interesting. Far from a domain expert in materials science, but I've worked professionally in this area. What's your method for identifying valid "novel" compounds? Certainly, anything actually novel has been included in the models' training set already, unless you're doing a CASP-like coordinated blind test...right?

As an aside, the "Fable lies and cheats" section made me laugh -- have encountered this same failure mode, albeit for much simpler models. Polymerization (I realize this is not exactly the same thing) is an unbounded 3D playing field for constraint escape. You can try to put an additional constraint on polymerization, but then it will just make minor variations on the monomers...

Sure, sure. So now we have the worst of all possible worlds: an expensive interview process where the candidate has to be onsite, and a stupid evaluation mechanism that doesn't at all reflect real-world job requirements. And in exchange for the inconvenience of an in-person interview, the candidate gets treated like a number and is met with a stupid memorization gauntlet that tests a skill that quite literally will never be used again now that we have AI. Sounds great.

Man, there are times when I really hate this industry.

> He did so. He aced the test. He got the job with a huge offer. He agreed that the ROI was out of the park!

...and everybody clapped?

I have news for you: passing an online leetcode is now so easy that an AI can do it. If you think you're catching the cheaters, you're just wrong.

Leetcode was always a stupid test of competence, but it was cheap and it had a high true-negative rate, which filtered the riff-raff. That isn't true anymore.

> To me, the difficult parts of software engineering are in the design. That's where you're defining the problem you're trying to solve, picking what technology to build upon, what kind of architecture you're planning on implementing, ruling out what does/doesn't make sense, taking input from stakeholders, and finding the balance between price and performance.

This is exactly why I became a PM [1]. I wanted to do "harder" things that had bigger scope than I could do as "just an engineer", and the kind of roles that offer that scope within engineering end up being thin on the ground. Unfortunately, what I've generally found is that being a technical PM is a good way to have people try to knife you constantly. Most PMs are technically illiterate MBA types, and carry a deep grudge against anyone who gives the technology any consideration. They're also good at politics, and want you to die. On the other side of the coin, because most PMs are technically illiterate, engineers often instinctively reject you like a tumor. To overcome this, you need to appeal to the technology, which puts more of a target on your back from the other PMs...

Anyway, my point is that this is all very sad, because you're describing is exactly what a PM does -- but with technical competence. And because of the way the industry is structured, the people who can do that either get shuffled into people management, or reach a rapid career ceiling when so-called "architecture" roles aren't available.

There's this pervasive myth that you cannot be good at technology while also prioritizing the business needs, but usually what this really means is that the "business types" bias in one direction exclusively, and treat the technology as the enemy. It's why so many companies are AI-maxxing now -- the promise of replacing the expensive nerds is the eternal flame for the MBA.

[1] well, that, plus after being a founder, people were suddenly willing to hire me to be a PM...

All of these things are strong covariates, which is something the online hive-mind struggles with. It's not intelligence OR wealth OR hard work OR $thing, it's all of the above, with some weighting function that spits out high values for all factors.

And since IQ is just a proxy for intelligence (however good), it also correlates to some greater or lesser degree with the other factors.

The benchmark article we're replying to shows that Grok token usage is at least on par with the latest OpenAI models [1], and significantly cheaper per token:

https://artificialanalysis.ai/models/grok-4-6#token-use

So depending on how you want to define "token efficiency", Grok is either tied with OpenAI, or in the lead.

[1] Though I grant that 4.6 appears to be wordier, on the order of Terra max.

Not that I know of. AA's token use metrics (mentioned in this article) are indicative, however. They say explicitly here that the Grok models are notably token efficient. This is my experience.

Yeah, I know, but "credit" translates differently because the models bill at different rates, which gets turned into "multipliers" (or at least, it did).

Have they converted entirely to transparent API rates + base allocation now? One of the reasons I left was that if I was going to be billed at API rates anyway, I'd just rather use the APIs. The value proposition still sucks for individuals now, when the other major providers are bundling at below-API rates.

If they didn’t constantly reset, they’d be about the same as Anthropic.

Right now, I find that Grok offers better value, uses fewer tokens per turn, and makes better code. I haven’t tried Cursor because I don’t want to change editors again, but maybe I should try it…

What’s the multiplier? GH copilot nerfed their product so badly that I unsubscribed.

> and i didn't mean to come off as too harsh. but i imagined someone reading your comment and being deeply discouraged (as i might have been, many years younger) when in reality tokyo has possibly one of the best and most accessible subway systems in the world.

Fair point. And I should apologize for being a bit snarky in response -- I get prickly when accused of "disinformation".

I admit that I'm used to the US ADA version of things, and that colors my views. But yeah, by the standards of the rest of the world, that's probably a pretty high bar.

> (and don't get me started on NYC, which is just wheelchair-hostile.)

Oh, definitely.

Not publicly. Send me an email.

> if there's a way in, and a way out, i consider that a win. will i lose maybe 5 minutes because the accessible entrance is not in the most convenient location that day? perhaps. but it's just that - an inconvenience.

I get that, and I also don't think they need to plop an elevator at every staircase, but there are so many stations where they could have ramps or some other minor improvement to make more exits accessible. And if the one path drops you on the wrong side of the tracks from your destination, that's often more than a five-minute problem -- my own home station has one elevator on the north side of the tracks, but if your destination is on the south, you must exit, cross a busy intersection, and cross the tracks at grade.

> on the staff support: i don't really see any problem here at all. in fact, i think the service that the staff provides is truly excellent.

I agree. I just made a similar comment on your other response. Compared to many places in the world, they're at least trying.

> I don't think i've ever seen a station with only one elevator, it's always at least one platform-to-ticket-gate, and one ticket-gate-to-ground (maybe that's what you meant).

Well, first, I'm talking not just about the subway, but all rail lines, which is a higher bar.

Setting that aside, what I meant was what you've inferred -- there's typically one elevator path that will get you out of an "accessible" station. Yes, that (usually) requires one elevator per platform to work at all, but it still leaves most exits inaccessible in the average case.

I started this subthread only by saying that Tokyo is not particularly wheelchair friendly. If you use one, I don't know how you could possibly dispute this. It's a land of staircases, and I've been repeatedly foiled just trying to move heavy luggage.

I grant you that in Japan they will usually bend over backwards to make something work, but I really feel for disabled folks here.

> I'd be interested learning how you've come up with that - it's disinformation. i'm speaking from extensive, direct experience.

I live here and have eyes. But tell me more about "disinformation."

Tokyo metro's standard for accessibility is to have one route that is accessible from street level to the platform. That's fine, but it's a far cry from every entrance being accessible, or even just ensuring that you don't require staff support to make it through that route. Consider Takadanobaba [1] -- a major transit point that I know well. Yes, it's technically "one route", but just taking into account subway station exits (and not Seibu or JR line transfers, which are harder), it qualifies only because it has a single elevator that goes to ground. Literally every other exit is inaccessible.

[1] https://www.tokyometro.jp/lang_en/station/takadanobaba/index...

Not even close. They're better than they used to be, and the major stations that tourists go to are mostly kinda-sorta OK, but the median station has 1 elevator if you're lucky, and probably around the same number of accessible entrances.

...and we haven't even discussed the fact that riding the train with a wheelchair typically involves reservation of a staff member on each end of the journey, to help you get on and off.

I had to laugh when they put a quote from a wheelchair user at the top of the article. Japan is many good things, but friendly to wheelchair users is not one of them.

As western tourists always seem to find out when they land with their gigantic suitcases, many Tokyo train stations require stairs to enter or exit.

It's amazing that they bothered to put together a slide summarizing their experience for a pitch deck.

Nothing. If the next person comes in to vote after you (which is easy enough to avoid if you use active voter lists), they might issue a provisional ballot. How do they know which voter is lying? They don’t check ID, and the other ballot isn’t associated with a name.

You have to be completely ideologically blinkered not to see a problem here.