Theo Sanderson
@theo.io
researching viral pathogen genomics at @lshtm.bsky.social | theo.io / sandersonlab.org
The summary sounds about right to me, and is also the message I got from the article
October 8, 2025 at 10:36 AM
The summary sounds about right to me, and is also the message I got from the article
do wonder if this could ultimately doom short-form video 🙏
October 7, 2025 at 10:39 PM
do wonder if this could ultimately doom short-form video 🙏
In "thinking mode" it generates a chain of "thoughts" [words] before answering "thinking through" the options so the response takes a few seconds to start but makes the right moves
October 4, 2025 at 11:43 PM
In "thinking mode" it generates a chain of "thoughts" [words] before answering "thinking through" the options so the response takes a few seconds to start but makes the right moves
(not intended as serious analysis!)
October 3, 2025 at 6:52 PM
(not intended as serious analysis!)
these temu true wireless earbuds have remarkable consumer surplus
August 21, 2025 at 4:33 PM
these temu true wireless earbuds have remarkable consumer surplus
yep, here I used the open-source model (so significantly less good) in order to be confident it wasn't running any code behind the scenes. last screenshots are sections of the "reasoning" trace [which is hidden in chatgpt]
August 13, 2025 at 1:13 PM
yep, here I used the open-source model (so significantly less good) in order to be confident it wasn't running any code behind the scenes. last screenshots are sections of the "reasoning" trace [which is hidden in chatgpt]
In the first layer of the network, the model needs to encode each amino acid into a numerical space. I constrained that space to be two-dimensional and what one ended up getting was an embedding a bit like the plot below - with axes representing polarity and size.
August 13, 2025 at 11:58 AM
In the first layer of the network, the model needs to encode each amino acid into a numerical space. I constrained that space to be two-dimensional and what one ended up getting was an embedding a bit like the plot below - with axes representing polarity and size.
Claude Code has got a lot more likely to refuse basic bioinformatic tasks recently (presumably due to dual-use concerns)
August 12, 2025 at 3:31 PM
Claude Code has got a lot more likely to refuse basic bioinformatic tasks recently (presumably due to dual-use concerns)
(made with chatgpt, naturally)
August 8, 2025 at 6:09 PM
(made with chatgpt, naturally)
There are more graphs if you follow the link
August 7, 2025 at 10:40 PM
There are more graphs if you follow the link
If you're talking a simple calculator like this, likely not?
August 7, 2025 at 9:56 PM
If you're talking a simple calculator like this, likely not?
sure one can use whatever analogy one wants, but it has interesting abilities
August 7, 2025 at 9:44 PM
sure one can use whatever analogy one wants, but it has interesting abilities
August 7, 2025 at 9:29 PM
Post an unusual sign if you feel like it
June 28, 2025 at 3:09 PM
Post an unusual sign if you feel like it
the daleks can't quite climb the stairs yet
June 3, 2025 at 10:18 PM
the daleks can't quite climb the stairs yet
so codex now has optional full internet access but still no way to terminate a job once started afaict 😓
June 3, 2025 at 9:08 PM
so codex now has optional full internet access but still no way to terminate a job once started afaict 😓
if asked to "write the opening of a thriller novel set in London" (or many variants on that), Claude [both Opus and Sonnet] invariably gives me one starring one "Detective Inspector Sarah Chen"
May 25, 2025 at 2:29 PM
if asked to "write the opening of a thriller novel set in London" (or many variants on that), Claude [both Opus and Sonnet] invariably gives me one starring one "Detective Inspector Sarah Chen"
![[tweet from anthropic person]
Initiative: Be careful about telling Opus to ‘be bold’ or ‘take initiative’ when you’ve given it access to real-world-facing tools. It tends a bit in that direction already, and can be easily nudged into really Getting Things Done.
If it thinks you’re doing something egregiously immoral, for example, like faking data in a pharmaceutical trial, it will use command-line tools to contact the press, contact regulators, try to lock you out of the relevant systems, or all of the above.](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:727oeq2js3j7ukz3tndao4sd/bafkreic774e2acnn6wmed3y6zag2qc3hdj7v3tkd5bjthfvu6q3abnzx3m@jpeg)