iiinova — where your digital self keeps evolving

Josh Blake @joshuablake.co.uk · 18/09/2026

This toy model shows the simple version of the argument, with many caveats. Not endorsing it but worth reading. www.forethought.org/research/how...

forethought.org

How quick and big would a software intelligence explosion be?

Forethought paper modeling software intelligence explosion: 60% chance of >3 years progress in <1 year.

100

Josh Blake @joshuablake.co.uk · 04/08/2026

They should at least care about durability of non-consumable goods. However, I think online shopping is basically the classic market for lemons. Reviews should solve this but don't. My best guess is reviewer standards aren't well calibrated.

110

Josh Blake @joshuablake.co.uk · 02/08/2026

Extremely useful, thank you!!

000

Josh Blake @joshuablake.co.uk · 01/08/2026

In practice, what do you do in this situation? Can you just jump on the next train and explain to a ticket inspector what happened or do you need some proof? And how much flexibility is there: do you need to still take the same route or if there's a reasonable quicker alternative are you ok?

200

Josh Blake @joshuablake.co.uk · 25/07/2026

The confidence intervals overlap massively. Suggests that reversal could just be noise.

010

Josh Blake @joshuablake.co.uk · 24/04/2026

Based on that, administering healthcare is about 15% of 20% = 3% increase in GDP. Moving UK from 51st to 50th. en.wikipedia.org/wiki/List_of...

en.wikipedia.org

List of U.S. states and territories by GDP - Wikipedia

010

Josh Blake @joshuablake.co.uk · 28/03/2026

> Some museums think they would actually end up in a worse position Why can't these museums just choose not to charge entrance fees?

010

Josh Blake @joshuablake.co.uk · 09/01/2026

Great study! Does this have implications for the effectiveness of ventilation or germicidal UV at preventing transmission?

010

Josh Blake @joshuablake.co.uk · 13/09/2025

Do the safety concerns here also extend to the built in web search tools, if you've provided sensitive information in your prompt?

110

Josh Blake @joshuablake.co.uk · 30/08/2025

This is only employment income right? Don't the very highest earners get most of their income from capital gains?

100

Josh Blake @joshuablake.co.uk · 29/04/2025

Supposedly fixed simonwillison.net/2025/Apr/29/...

000

Josh Blake @joshuablake.co.uk · 09/04/2025

First one makes searching easier IMO

100

Josh Blake @joshuablake.co.uk · 23/03/2025

The 80/20 Rule: if you've solved 80% of the problem, you've spent 20% of the resources required.

130

Josh Blake @joshuablake.co.uk · 14/03/2025

Defer to CEA. If they accept you they think it's worth having you there. If you're worried about wasting others money then you can buy the full cost ticket.

020

Josh Blake @joshuablake.co.uk · 06/02/2025

(I think I might be too low context on what argument is going on here too usefully engage so feel free to ignore)

020

Josh Blake @joshuablake.co.uk · 06/02/2025

I generally don't support making bad arguments to counter bad arguments

110

Josh Blake @joshuablake.co.uk · 06/02/2025

Argument proves too much IMO, basically lets you disregard all thought experiments where you don't like the outcome. Which I don't think you or Scott would agree with. (It's not really an argument, it's more rhetoric anyway.)

230

Josh Blake @joshuablake.co.uk · 31/01/2025

This is because of my high respect for Rethink Priorities (which Peter was co-CEO of until recently), his forecasting track record, and the quality of a handful of old blog posts I've read by him. Not the current post which is mostly "coming soon".

000

Josh Blake @joshuablake.co.uk · 31/01/2025

Omg I've never been so excited by a Substack

151

Josh Blake @joshuablake.co.uk · 27/01/2025

Do you have a good initial prompt to get an interesting conversation rather than an almost lecture each time you ask something?

110

Josh Blake @joshuablake.co.uk · 24/01/2025

@helenbranswell.bsky.social. But honestly, there's very little signal and most information is noise (unless you're trying to work out detailed policy).

010

Josh Blake @joshuablake.co.uk · 05/01/2025

What do power calculations provide that confidence intervals don't? In both cases, you're asking how big an effect is compatible with the observed data.

000

Josh Blake @joshuablake.co.uk · 05/01/2025

This search strategy kinda reminds me of how Prolog works. But you've now got a massive fact base (training data) and a fuzzier goal (Raters)

000

Reposted by Josh Blake

Simon Willison @simonwillison.net · 04/01/2025

Here's a question I have NEVER managed to get a straight answer to: if I paste an internal company strategy document into ChatGPT or similar to ask for a summary, what are the chances that a competitor in six months time can ask the model about my strategy and get it to leak internal secrets?

7489

Josh Blake @joshuablake.co.uk · 30/12/2024

Then there's the same question about the right combination of mutations arising though right? Somewhat complicated by the mutations possibly happening in non-human hosts.

200

Josh Blake @joshuablake.co.uk · 30/12/2024

Can we apply this to a question roughly "at what point have we seen so many dairy workers infected without (enough to be detected) onwards spread that we should conclude this is so unlikely to cause an epidemic that we should focus efforts elsewhere"?

100

Josh Blake @joshuablake.co.uk · 30/12/2024

2. More human infections (/increasing rate) is more rolls of the dice BUT each infection reduces our posterior probability of any infection becoming a pandemic. How do these shake out in overall update of risk? Probably not an original thought - maybe there's work on this already. 2/2

130

Josh Blake @joshuablake.co.uk · 30/12/2024

I do sometimes wonder if we're looking in the wrong places. 1. This paper (screenshotted) suggesting that purely zoonotic (no h2h transmission) don't become epidemic, it's the ones with small outbreaks that do. academic.oup.com/mbe/article/... 1/2

Paper screenshot with text: "The phylogenetic pattern is consistent with epidemiological experience of human RNA viruses going back over 100 yr. There have been multiple examples of the emergence of new human-transmissible viruses from nonhuman reservoirs (including HIV-1 and SARS-CoV-2) and several examples of outbreak (L3 by definition) viruses becoming epidemic (L4) viruses (including Ebola virus and Chikungunya virus). In contrast, there have been no clear examples of strictly zoonotic (L2) viruses becoming human transmissible (though the possibility has been raised as a concern, especially for the avian influenzas; Woolhouse et al. 2016)."

250

Josh Blake @joshuablake.co.uk · 26/12/2024

What's does Target have that Tesco Extra or massive Sainsbury's don't?

110

Josh Blake @joshuablake.co.uk · 23/12/2024

Should have tagged the author @binarybits.bsky.social as a consistently insightful voice on AI, and probably the best writer on self-driving cars.

020

Josh Blake @joshuablake.co.uk · 23/12/2024

If you're not obsessively following AI, strong recommend below for implications of the latest models focusing on a new type of training and compute paradigm. Clearly explains what's different and the implications. Worth the subscription just for this article IMO. open.substack.com/pub/understa...

open.substack.com

It's the end of pretraining as we know it

LLM training techniques are evolving, and that has huge implications.

271

Reposted by Josh Blake

Infectious Disease Tracker @hmpxvt.bsky.social · 20/12/2024

WHO declares Rwanda’s Marburg virus outbreak over. 65 cases and 15 deaths were reported.

1679

Josh Blake @joshuablake.co.uk · 10/12/2024

Is there any reason to suspect H5 here?

000

Josh Blake @joshuablake.co.uk · 09/12/2024

How confident are you about the signalling and negotiating power claims? These aren't obviously correct to me, especially the one about group randomisation as well as the interaction with a FPTP voting system.

100

Josh Blake @joshuablake.co.uk · 09/12/2024

Somewhat, but not fully because their output is too dependent on the prompt. This gives you degrees of freedom. You'd want to somehow show that it was the first prompt you tried. (Also issues around possible bias of the LLM itself.)

120

Josh Blake @joshuablake.co.uk · 07/12/2024

For example, the central limit theorem (limit as sample size n->infinity) describes behaviour well for large n and makes calculating confidence intervals possible without assuming a full model

010

Josh Blake @joshuablake.co.uk · 07/12/2024

I think he means the behaviour of a system or function as a parameter x tends to infinity (or zero) approximates the situation when x is very large (or zero). This is often useful because the behaviour in the limit is simpler.

120

Reposted by Josh Blake

Paul Novosad @novosad.bsky.social · 06/12/2024

Fascinating paper on where 6000 global elites went to college. Billionaires, CEOs, heads of state, central bankers, etc. In a word: Harvard. Fully 10% of global elites went to Harvard. Elite US schools are over-represented (23% IvyPlus), but nobody comes close to Harvard. 🧵

37709296

Reposted by Josh Blake

Meaghan Kall @kallmemeg.bsky.social · 02/12/2024

I spoke to @gavinfreeguard.bsky.social for a little glimpse behind the (imperfect but real-life) data that informed the UK pandemic response. Have we improved our infectious disease data systems since March 2020? Yes and no…

286

Josh Blake @joshuablake.co.uk · 04/12/2024

Hmm interesting, thank you for sharing

000

Josh Blake @joshuablake.co.uk · 03/12/2024

Cheese grater. You never check you have a cheese grater until you want to grate cheese. Then it's too late.

010

Josh Blake @joshuablake.co.uk · 02/12/2024

@andymatuschak.org big fan of your work and ideas! I was wondering if you've ever written what tools you use for reading and writing inbox, including incorporating spaced repetition? I love the ideas in your working notes but struggling to see how to implement them.

110

Josh Blake @joshuablake.co.uk · 02/12/2024

I want to keep better track of frontier AI from a range of perspectives. Recommendations for Substacks (preferred), blogs, or people to follow (preferred not) which provides news in a shortish and accesible form, without ridiculous hype, ideally with a focus on putting developments into context.

000

Josh Blake @joshuablake.co.uk · 02/12/2024

I strongly believe in avoiding nuclear winter* *This makes no judgement about the possibility or likelihood of nuclear winter

010