Built Open Artificial Pancreas (@OpenAPS) with @DanaMLewis. Day job @Netskope. DMs are open or email [email protected]

Seattle area
Joined May 2011
Another open-source tool powered by @OpenAI #Codex: auto-docstring (github.com/scottleibrand/cod…) writes and adds Python (PEP 257) docstrings summarizing what each of your functions does, what arguments it takes, and what it returns. How/where would *you* use this kind of capability?
And it works! github.com/scottleibrand/cod… is a new open-source tool powered by @OpenAI #Codex that writes and adds inline comments explaining whatever code you give it. What code would *you* like comments added to? Reply with a link to it on GitHub, and I'll give it a try.
4
16
4
80
Scott Leibrand retweeted
This looks like a permanent shift, and a really big one too. From 30% to almost 80% just like that. A change on this scale is bound to have all sorts of interesting consequences.
71
440
47
2,795
Scott Leibrand retweeted
I think this is an excellent take on a subtle but important problem.
How can uncertainty in uncertainty be handled in a consistent way? - a puzzle: (1) Suppose you think a weighted coin has a ~60% chance of landing on heads (because you flipped it 1 million times and it landed on heads 60% of the time).
Show this thread
2
1
0
8
*~*VARIANT UPDATE*~* I took a little break from these but back with an important new report @ukhsa has designated AY4.2 a VUI VUI-21OCT-01 Let’s break it down 🧵 ⤵️ Technical Briefing 26 - AY.4.2 assets.publishing.service.go…
36
422
102
949
Show this thread
Scott Leibrand retweeted
do I think the global supply chain backup will be solved by This One Trick? No. The question is why aren’t those directly in charge trying everything? my view: because lots of ideas require someone ELSE to say OK so you get stuck. Anyone who’s worked in a big org understands
7
28
4
776
Show this thread
Some late-night laying-awake thoughts inspired by this Truthful AI paper (h/t @anderssandberg). One key capability we need, for Truthful AI and more generally, is the ability for an ML model to evaluate purported statements of fact for likely accuracy.
New paper on truthful AI! We introduce a definition of “lying” for AI We explore how to train truthful ML models We propose institutions to support *standards* for truthful AI We weigh costs/benefits (economy + AI Safety) (w/ coauthors at Oxford & OpenAI) arxiv.org/abs/2110.06674
Show this thread
1
1
1
2
So maybe we should start developing a language model that can evaluate the truthfulness of statements of fact, and thereby limit itself to making only such statements.
1
0
0
0
Where to start? Here’s one idea: If a large language model is suitably fine-tuned or trained on the edit history of Wikipedia, it should be able to predict which statements survive, which are edited, and which are cut.
1
0
0
0
Trained in this way, it could evaluate the quality of a novel statement, which would predict some combination of its truthfulness and quality (relevance, coherence, etc.).
1
0
0
0
Is this doable today? Maybe? For today’s language models like GPT-3, it might be best to start with the simplified English version of Wikipedia, on the assumption that would be more likely to be within its capabilities to “understand”.
1
0
0
0
How would I try something like that? First, I’d need to figure out how to download the full edit history for Wikipedia pages, and whether to use that as-is for fine tuning or whether to process that into a form that could be used without fine-tuning via prompt engineering.
0
0
0
0
Scott Leibrand retweeted
One way that AI could make our economy explode: cold-takes.com/the-duplicato… See also:
1) Most economists think GDP growth will be slower the coming decades, due to recent trends in productivity and population size. Afaik none predict an acceleration of growth, but I think there's a reasonable chance of this (say 10-30% within 50 years).
Show this thread
0
1
0
1
Scott Leibrand retweeted
.@HoldenKarnofsky of GiveWell and Open Phil fame has a new blog and it's as good as you would expect: "...growth [must be] near its historical high point. There aren't enough atoms in the galaxy to sustain today's growth for even another 10,000 years." cold-takes.com/this-cant-go-…
5
19
2
73
Show this thread
We have budget remaining from Convening The Center that we'd like to use to provide nine people with $1,000 USD internet scholarships. Can you help us find the right people who could use internet scholarships to do good in healthcare? Details: bit.ly/CTC-Internet-Scholars…
3
13
2
7
Show this thread
Scott Leibrand retweeted
Pfizer continues to impress: 2 Pfizer doses of 10μg each (doses for adults are 30μg) has efficacy of ~91% against symptomatic COVID in children 5-11. Study was done at a time during which Delta variant was predominant fda.gov/media/153409/d
0
1
0
2
Scott Leibrand retweeted
Tether is used to buy Bitcoin, which is used as collateral to borrow Tether, which is used to buy more Bitcoin when prices drop to pump the price, which creates fomo so that people will buy more Bitcoin.
New from @ben_mckenzie and me: we took a close look at Tether, the unofficial (and definitely unregulated) central bank for crypto's wild form of casino capitalism. slate.com/technology/2021/10…
Show this thread
4
17
1
53