1. The thing about science that these jokers don't understand is that science cannot be vibe-coded.
Whatever its flaws, the point with vibe coding is that you're trying to quickly make something that sorta works, where you can immediately sorta see if it sorta works and then sorta use it.
Last term I tried an experiment: I walked into my Tech and Design Ethics class, admitted that I had *no idea* what to do about ChatGPT - so I would let them figure it out.
As in: their first project was to decide and write the ChatGPT policy for the class.
Here's what happened:
Physicist Leo Szilard, in a short science fiction story from 1948, describing how to retard science by making the funding application longer and harder than the proposed research - now called the ‘Szilard point’
This article suggests the possibility of training AI on pre-existing proposals and their review reports, scores and related decisions. I.e. training them on a system which is known to preferentially award larger amounts of money to white men. What could possibly go wrong?
Basic principles for the future of the car in cities:
✅ Fewer
✅ Smaller
✅ Slower
✅ Cleaner
This can be achieved with:
✅ Parking controls
✅ LTNs
✅ Road User Pricing
✅ Pedestrianisation
✅ Default 20mph roads
✅ Improved public transport
✅ Segregated cycling
Oh no! It was great to meet you up here in Edinburgh. Such a shame the trains are so often a mess.
Quite common, and I prefer it. If your calendar is up to date it saves a lot of emails/messages between people. And you can always move it back in the same way!
This is a good point! What frustrates me most is that these models are not actually trained on research papers. Copyright and environmental issues aside, a ChatGPT-like model trained on all scientific literature and books could be a really useful thing.
The sort of good quality answers that you posted build confidence, but then suddenly you are faced with something that doesn't quite make sense. Only through deep personal knowledge can you notice it. My fear is that the majority of users will never know. That makes me uneasy.
No, I think it is a perfectly good answer. But I agree with @jacopobertolotti.com that it is Wikipedia level. That is not a problem, and the fact that you get a specific direct answer is really useful. My issue is for the more subtle edge cases, where genuinely deep understanding is needed.
if you see this post, your actions are:
- if you have a spare buck, give it to Wikipedia, then repost this
- if you don't have a spare buck, just repost
your action is mandatory for the world's best source of information to survive
I don't have an explicit example off the top of my head, but so far (even with GPT-5 thinking) I have found many subtle errors when discussing physics. I'll try to remember to post one when it next occurs.
That is the main problem. When humans don't know something or start to BS, it is obvious. When chatbots don't "know" something it is often indistinguishable from when they are correct, unless you already know the answer. So we must use them as ignorant, but great, language manipulation tools.
Chatbots — LLMs — do not know facts and are not designed to be able to accurately answer factual questions. They are designed to find and mimic patterns of words, probabilistically. When they’re “right” it’s because correct things are often written down, so those patterns are frequent. That’s all.
It astonishes how often the answers are wrong (even pro-level tools). And when wrong, it is often in subtle and hard to notice ways. When constrained appropriately, it can be very useful. But it is exceedingly scary how many people take the authoritive sounding and articulate answers as fact. (2/2)
I disagree. I've been trying to make better use of this technology (partly motivated by you!) and made significantly more progress with the perspective that these things do not think at all and just complete maximal likelihood sentences. It is then a great tool for text or code wrangling. (1/2)
Code is Just, all three seasons, as heavily requested, is finally out in a compiled PDF and ePub for easy reading. Price is whatever is reasonable for you, including zero. 1720541891659.gumroad.com/l/qkali
"Matlab is the Visual Basic 6 of technical computing." #HeardAtJuliaCon #JuliaCon #JuliaLang
I'm curious what your use cases are (within academia), especially as you seem quite optimistic about AI use. I mostly try out some AI tools every few months, have a brief wave of optimism, but then give up and go back to my old ways after finding the output not as good as I thought.
If you love supercontinuum (and who doesn't!) check out our forward-looking Perspective in EPL which is free to read. Thanks @jctravs.bsky.social Thibaut Sylvestre, Alex Heidt, Roy Taylor, Goery Genty iopscience.iop.org/article/10.1...
If you are applying to a PhD, don't use an LLM in composing your proposal.
If you are doing a PhD, don't use an LLM to do the writing and reading and thinking for you.
Sorry, I am going to die on this hill.
Post-deadline PD-2.3 18:05 ICM - Room 13b Extreme Soliton Dynamics for Terawatt-Scale Optical Attosecond Pulses and 30 GW-Scale Sub-3 fs Far-ultraviolet Pulses Nikoleta Kotsina, Michael Heynck, Joleik Nordmann, Martin Gebhardt, Teodora Grigorova, Christian Brahms, and John C. Travers Heriot-Watt University, Edinburgh, United Kingdom We present the extreme soliton (XSOL) beamline, which generates terawatt-scale sub-femtosecond self-compressed visible-infrared pulses as well as sub-3 fs far-ultraviolet pulses with energy exceeding 170 µJ at 240 nm and wavelength tunability down to 140 nm.
Visit my group's post-deadline talk at CLEO/Europe-EQEC to find out what's been consuming an enormous amount of our time, energy and money(!) lately.
Nikoleta Kotsina will present TW-scale optical attosecond pulses and sub-3 fs DUV pulses with more than 170 µJ of energy!!
🧪💡⚛️😎
Congratulations to @nickkarpowicz.bsky.social and co-authors! This is a major contribution to our field.
I am still digesting the details, but this is a wonderful piece of work!
🎉New paper! doi.org/10.1364/OE.5... 📝
We look at how different numbers of capillaries affect the guidance properties of hollow core fibres, enabling the guidance of ~ 50 modes in the near-infrared. Some cool fibres in here that guide the visible and ultraviolet #science #optics #photonics
If you burn down a forest, you don't miss out on lumber for just that season. You have to replant all the trees, nurture them, and wait for them to grow.
The science and research budget cuts happening now are wanton, senseless arson. Recovery, if it ever happens, will take generations.
Why are my lasers so expensive?
Our optica paper, "Tunable megawatt-scale sub-20 fs visible pulses from a fiber laser source", is out!
Happy to have been a part of this during my time at Heriot-Watt
opg.optica.org/optica/fullt...
Announcing a PhD position starting October 2025, on Applications of Machine Learning in Ultrafast Nonlinear Fibre Optics. The advertisement will be out soon, but my email is easy to find, so please send me a CV if interested. Experience in experimental & numerical photonics highly desirable!
Even accepting the premise that AI produces useful writing (which no one should), using AI in education is like using a forklift at the gym. The weights do not actually need to be moved from place to place. That is not the work. The work is what happens within you.
The @uniofbath.bsky.social Optical Fibres group is hiring! They are looking for a Professor and two Lecturers to join. If bespoke optical fibres could benefit your research, this is an opportunity. More info: www.bath.ac.uk/research-gro...