Hah! Well found. I do recall hearing about another simulated vendor experiment (that also failed) but not actual dog-fooding. Looks like the big upgrade the wsj reported on was the secondary “seymour cash” 🙄 chatbot bolted on the side… the main chatbot was still claude v3.7, but maybe they’d prompted it harder and called that an upgrade.
I wonder if anthropic trialled that in house, and none of them were smart enough to break it, and that’s what lead to the external trial.
So, I’m taking this one with a pinch of salt, but it is entertaining: “We Let AI Run Our Office Vending Machine. It Lost Hundreds of Dollars.”
The whole exercise was clearly totally pointless and didn’t solve anything that needed solving (like every other “ai” project, i guess) but it does give a small but interesting window into the mindset of people who have only one shitty tool and are trying to make it do everything. Your chatbot is too easily lead astray? Use another chatbot to keep it in line! Honestly, I thought they were already doing this… I guess it was just to expensive or something, but now the price/desperation curves have intersected
Anthropic had already run into many of the same problems with Claudius internally so it created v2, powered by a better model, Sonnet 4.5. It also introduced a new AI boss: Seymour Cash, a separate CEO bot programmed to keep Claudius in line. So after a week, we were ready for the sequel.
Just one more chatbot, bro. Then prompt injection will become impossible. Just one more chatbot. I swear.
Anthropic and Andon said Claudius might have unraveled because its context window filled up. As more instructions, conversations and history piled in, the model had more to retain—making it easier to lose track of goals, priorities and guardrails. Graham also said the model used in the Claudius experiment has fewer guardrails than those deployed to Anthropic’s Claude users.
Sorry, I meant just one more guardrail. And another ten thousand tokens capacity in the context window. That’ll fix it forever.
So, curse you for making me check the actual source material (it was freely available online, which seems somehow heretical. I was intending to torrent it, and I’m almost disappointed I didn’t have to) and it seems I’m wrong here… the anconia copper mine in the gulch produced like a pound of copper, and they use mules for transporting stuff to and from the mine. The oil well is still suspicious, but it just gets glossed over.
I’m not prepared to read any more than that, so if there was anything else about automation in there I didn’t see it. I’d forgotten the sheer volume of baseless smug that libertarian literature exudes.
I can’t find any good sources right now, I’m absolutely not going to find a copy of that particular screed and look up the relevant bits, but I think the oil wells in galts gulch were automated, with self driving trucks, I think? It might also be implicit in the goods that are produced there, but that is drifting a bit into fanfic, I admit.
Anyway, between boston dynamics and the endless supply of rand fans on the internet, it’s very hard to research without reading the damn thing.
If I find a new source, I’ll report back.
They had robots in galts gulch, which means that all businesses need them. If you aren’t randmaxxing 24/7, can you really call yourself a technological visionary at the vanguard of the libertarian master race?
Of all the environments that you might want to rearrange to facilitate non-humanoid labour, surely warehouses are the easiest. There’s even a whole load of pre-existing automated warehousing stuff out there already. Wheels, castors, conveyors, scissor lifts… most humans don’t have these things, and they’re ideal for moving box-like things around.
Industrialisation and previous waves of automation have lead to workplaces being rearranged to make things cheaper or faster to make, or both, but somehow the robot companies think this won’t happen again? The only thing that seems to be different this time around, is that llms have shown that the world’s c-suites are packed with deeply gullible people and we now have a load of new technology for manipulating and exploiting them.
Much of the content of mythical man month is still depressingly relevant, especially in conjunction with brooks’ later stuff like no silver bullets. A lot of senior tech management either never read it, or read it so long ago that they forgot the relevant points beyond the title.
It’s interesting that clausewitz doesn’t appear in lw discussions. That seems like a big point in favour of his writing.
A second post on software project management in a week, this one from deadsimpletech: failed software projects are strategic failures.
A window into another it disaster I wasn’t aware of, but clearly there is no shortage of those. An australian one this time.
And of course, without having at least some of that expertise in-house, they found themselves completely unable to identify that Accenture was either incompetent, actively gouging them or both.
(spoiler alert, it was both)
Interesting mention of clausewitz in the context of management, which gives me pause a bit because techbros famously love the “art of war”, probably because sun tzu was patiently explaining obvious things to idiots and that works well on them. “On war” might be a better text, I guess.
I chose to believe the evidence in front of my eyes over the talking points about how SpaceX was decades ahead of everyone else, SpaceX is a leader in cheap reusable spacecraft, iterative development is great, etc.
I suspect that part of the problem is that there is company in there that’s doing a pretty amazing job of reusable rocketry at lower prices than everyone else under the guidance of a skilled leader who is also technically competent, except that leader is gwynne shotwell who is ultimately beholden to an idiot manchild who wants his flying cybertruck just the way he imagines it, and cannot be gainsayed.
Bleugh, I’ve been using crucial ram and flash for a hell of a long time, and they’ve always been high quality and reasonably priced. I dislike having to find new manufacturers who don’t suck, especially as the answer seems to be increasingly “lol, there are no such companies”.
Thanks to the ongoing situation in the us, it doesn’t look like the ai bubble is going to pop soon, but I can definitely see it causing more damage like this before the event.
For a lot of this stuff at the larger end of the scale, the problem mostly seems to be a complete lack of accountability and consequences, combined with there being, like, four contractors capable of doing the work, with three giant accountancy firms able to audit the books.
Giant government projects always seem to be a disaster, be they construction, heathcare, IT, and no heads ever roll. Fujitsu was still getting contracts from the UK government even after it was clear they’d been covering up the absolute clusterfuck that was their post office system that resulted in people being driven to poverty and suicide.
At the smaller scale, well. “No warranty or fitness for any particular purpose” is the whole of the software industry outside of safety critical firmware sort of things. We have to expend an enormous amount of effort to get our products at work CE certified so we’re allowed to sell them, but the software that runs them? we can shovel that shit out of the door and no-one cares.
I’m not sure will ever escape “move fast and break things” this side of a civilisation-toppling catastrophe. Which we might get.
Sunday afternoon slack period entertainment: image generation prompt “engineers” getting all wound up about people stealing their prompts and styles and passing off hard work as their own. Who would do such a thing?
https://bsky.app/profile/arif.bsky.social/post/3mahhivnmnk23
Ahh, sweet schadenfreude.
I wonder if they’ve considered that it might actually be possible to get a reasonable imitation of their original prompt by using an llm to describe the generated image, and just tack on “more photorealistic, bigger boobies” to win at imagine generation.