tomjohnneill

Now you've successfully trained the next model not to put anything suspicious in its chain of thought. And unfortunately that's exactly what OpenAI and Anthropic seem to be doing (sometimes accidentally)

Did the agents you built work? I think we're doing something similar in the UK and intrigued to know how it went for you (https://www.plinth.org.uk)

I'd be very interested to hear more about how the non-for-profit agents platform went/anything else you learnt.

And it's ASML more than TSMC.

Or maybe there's a highly profitable role for all the different parts of the value chain.

Plinth (https://plinth.org.uk/) | London (UK) | Senior Engineer (mostly Frontend) | Onsite

At plinth, we've built a software/AI platform for charities to help them measure their impact and get more funding for the work they're doing. We work with both charities and funders (e.g. foundations, Government), providing a way to easily collect, visualise and report on client and impact data.

We're looking for an engineer with experience with React, NextJS etc, who wants to spend almost all their time heads down writing code, and not spending time in meetings.

To apply, try our puzzle: https://www.plinth.org.uk/puzzle

I think it's a case of tech bubble vs the rest of the world. Most people are not subscribing to the paid version of ChatGPT, but a lot of people who spend a lot of time with these things are.

It's the default for the free version of ChatGPT no? That's what the majority of people use.

I'm not sure. The figures I've seen suggest that GPT3 required 10x more energy to train than GPT2 (e.g. https://www.nnlabs.org/power-requirements-of-large-language-....), so I think a roughly 1-2 order of magnitude increase in energy usage from GPT2 to GPT3.5 makes sense.

I can definitely imagine they're not covering the amortised cost of the training with the cost per individual inference request. It seems less likely to me that they're making a significant loss on each subsequent request, but again no source from me on that either.

Looking a bit more into this, I found this paper: https://arxiv.org/pdf/2311.16863.pdf. It references a table saying that text generation uses 0.047 kWh per 1000 inferences, which is 1-2 orders of magnitude lower than my estimate. Though that is for GPT2, so possibly tracks to something roughly in the ~0.001 kWh per inference for GPT3.5.

That's fair about ignoring the training cost. I did write a bit more going into that in a follow up piece here: https://notfunatparties.substack.com/p/ai-is-good-for-the-pl...

Do you have any better sources for the power usage stats? It would be good to get a bit closer on that front. Having said that, even if the cost share is closer to 80%, that still puts it on par with a laptop for an average person.

Maybe decoding Herculaneum scrolls? https://en.wikipedia.org/wiki/Herculaneum_papyri#Virtual_unr...

It's been done (unsurprisingly): https://www.theaccessgroup.com/en-gb/products/mosaic/

That's actually not correct. Stripe invoicing charges 0.4 or 0.5% on top of the normal payment transaction fee (for either card payments or bank transfers). Essentially, you are paying a % fee for a nice PDF (plus some reconciliation/reporting etc)

They fixed that fairly recently.

I'm not sure the API docs are publicly accessible?

It's very likely organisations in the charity sector will need to run DBS checks, even just for looking at back-end systems/data.

Interesting to read those valuation benchmarks. I'd seen this article in SeedLegals previously https://seedlegals.com/resources/how-to-value-your-company/ that seem to be quite a lot lower.

It was written at the end of 2019, do you think it's just wrong? Or has it changed that much in the past 2 years? Or is it a different set of startups?

Paywalled, but: https://theathletic.com/2755968/2021/08/09/cox-is-heading-ab...

The main suggestion is that heading could be banned outside the 2 penalty boxes.

A UBI that doesn't apply to foreign migrants isn't exactly universal.

Also, the second order impact of oil money appreciating the currency and therefore the reducing the international competitiveness of non-oil industries definitely has a big impact. In contrast, in a UBI situation, GDP still depends on the productivity of the country, not just the wealth generated from a niche (in terms of employment) extractive sector.

Yeah it was statically hosted in the end (initially it wasn't quite static). When it was at its most popular, the total site size including JavaScript was around 70kb as far as I can remember, so it didn't come anywhere near as high as 70TB.

Part of my worry was that I messed up the caching initially so it was making API calls on the server for each request. Once I fixed that, there wasn't really anything to worry about.

It was basically a typo, but then enough people got slightly irritated by it that I started to enjoy having it that way.

It was a static site cached on Vercel. Not including the map (which is an iframe) the site was less than 100kb.

I wrote that up here: https://notfunatparties.substack.com/p/inside-a-viral-websit...

Provided you gain appropriate consent, there are definitely ways in which you can discover exactly how much the taste of these fluids vary from person to person.

Yeah I don't think anyone is ever going to build a better Goodreads from a bookmarking & automated recommendations perspective for exactly that reason. I think there is scope for some nice little niche tools for individuals to recommend books though. Whether that's the one I made in a weekend, seems doubtful, but I think there is a path there.

In hindsight, you're right. And obviously I didn't. However at the time, it was the uncertainty as a result of never having done something like this that made me worried.

For what it's worth (and I know this is getting extremely meta), but as it stands this writeup has had 21,963 visitors from Hacker News. (205 comments as I write this).

The point is that NextJS does create a static site. And Vercel can be just a static CDN, as it was for this.

Damn I wish you hadn't posted. Ah well.