Remote

Donal Fellows

@dkf@cyberplace.social

mastodon 4.7.3

Research Software Architect for Research Software Engineering at the University of Manchester.

Tcl programmer, but also many other techs too. Lots of software automation engineering these days.

I want real AGI, not fake stuff powered by overtrained LLMs.

0 Followers

0 Following

11 Posts

Joined March 04, 2025

Best Language:

Tcl

Best Project:

SpiNNaker

Donal Fellows @dkf@cyberplace.social

· 2mo ago

Replying to

@johncoxon@wandering.shop Nature does pretty well when you just leave it alone. Yes, there's a natural shifting of what plants do best over time, and for most of the UK, the real endpoint is forest. The sycamores invade my garden every year...

Donal Fellows @dkf@cyberplace.social

· 2mo ago

Replying to

@neilmadden@infosec.exchange Python does fairly well at tracking this sort of thing via PEPs. There might be some loss from the early stages, but that seems to be often true due to nobody realising at the time that it was important to keep it. Transitioning between VCSs is a key time when there's information loss. Or when the product gets moved between companies.

Donal Fellows @dkf@cyberplace.social

· 8mo ago

Replying to

@tauon @eloy On some platforms, you can use the Outlook application instead of the web app. Not that that's meaningfully better. I really miss being able to use Thunderbird.

Donal Fellows @dkf@cyberplace.social

· 4mo ago

Replying to

@CGM@mastodon.scot But I can't do anything about that myself; I'm on other projects (in atmospheric science and materials science) because they had the need for my skills and the money to pay for my work. Comes of being an internal consultant software engineer.

Donal Fellows @dkf@cyberplace.social

· 4mo ago

Normally, I don't share LinkedIn junk... but this time I'll make an exception:
https://www.linkedin.com/posts/congratulations-to-our-research-software-share-7459961130879193088-YKAp?utm_source=share&utm_medium=member_desktop&rcm=ACoAABiy5lEBsYvrxI1O5NkWVzsQjhsnkVw1EvE

linkedin.com

Donal Fellows @dkf@cyberplace.social

· 5mo ago

Replying to

@wyatt_h_knott@mstdn.social @quixoticgeek@social.v.st "Natural gas" is almost pure methane, which burns extremely cleanly and efficiently. Propane has a higher proportion of carbon, and lower of hydrogen, but also burns cleanly (with a somewhat different configuration of burner due to needing more oxygen per unit volume of gas).

Donal Fellows @dkf@cyberplace.social

· 7mo ago

@rayckeith@techhub.social Don't think of him as Batman. Think of him as more like Lex Luthor. He even looks the part.

Donal Fellows @dkf@cyberplace.social

· 7mo ago

They perform projections in a (very!) high-order smooth differentiable space. For problems which that describes, they work excellently. But when the problem doesn't fit that pattern, they're not so great. (The underlying math of hypersurfaces doesn't require differentiability, but the training algorithm guarantees it.)

Public LLMs have an awfully large training dataset. Much bigger than you think.

Donal Fellows @dkf@cyberplace.social

· 5mo ago

Replying to

#2 is a whole bunch of new functions, for example: set y [expr { $x1 * erf($x2) }]

Donal Fellows @dkf@cyberplace.social

· 4mo ago

Replying to

@CGM@mastodon.scot Exactly. It's a completely different technology direction, focused more around brain simulation. Turns out, if you simulate enough, you get something quite AI-like, except with very different limitations. I hope the team gets to investigate the key missing cell types (cortical pyramidal neurons, Purkinje neurons) as the result would have interesting applications.