Simon Brocklehurst retweetledi
@neilturkewitz This is why I believe LLMs to be incompatible with fair use doctrine
English
Simon Brocklehurst
984 posts
Simon Brocklehurst
@SMBrocklehurst
Personalized, Private Artificial Human-Like Intelligence. CEO AI startup, biotech exec, advanced software and robotics. Ex @UniOfOxford | Ex @Cambridge_Uni
Cambridge, England Katılım Haziran 2018
308 Takip Edilen143 Takipçiler
Simon Brocklehurst retweetledi
@rewtoetzi @MaxALittle @MartinJBCoulter It depends what you mean. None of that is taken into account *explicitly*. However, because it makes use of known native states of proteins, much of it is taken into account *implicitly*: that is, it bypasses the need to consider biological kinetic pathways of folding.
English
@SMBrocklehurst @MaxALittle @MartinJBCoulter How much is this taken into account? -
Protein folding dynamics in living cells is highly constrained, for example by chaperone proteins, enzymes, and other micro-structures near the ribosomes, to get the "correct", i.e. biologically useful folding, especially for large proteins.
English
@MaxALittle For sure, Tesla is taking the "get more training data and success is guaranteed" approach to perception. They seem have an at least partly symbolic approach to planning, though. It'll be interesting to see how far they get with their approach. My view is - it can't work as is.
English
@IntuitMachine I think they're concerned (rightly or wrongly) that the net effect of people flying in from Europe might be a significant increase in internal travel in the US, which could spread the Delta variant even faster across multiple regions of the US.
English
Can anyone tell me why the US is banning travel from Europe because of the Delta variant, yet the Delta variant is running wild in the Southern USA?
English
@WiringTheBrain When you say "encodings" do you mean "representations"? *ducks*
English
More progress towards properly holistic (dynamic, population-level) analyses of neural encodings 👍
srdjan ostojic@ostojic_srdjan
Distinguishing embedding (=linear) and intrinsic (=non-linear) dimensionality of neural activity: the published version of the opinion piece with @mjaztwit is now online (open access link): authors.elsevier.com/a/1dm0z3Q9h270…
English
@SMBrocklehurst Simon hahahahha. I never got a chance to ever play games on mine. Just my name and few bits flashing from side to side on the telly. Kids will never know our pain or the hours of fun we had.
English
RIP Clive Sinclair. Very sad news. His first computer, the ZX80, was the first computer I owned, and what I taught myself to program on. He inspired millions of kids to learn to program who went on to create entire industries. He made a dent in the universe. #CliveSinclair
English
@jonathan_mccrea @WiringTheBrain @smithbarryc @BBCRadio4 Defining things as "real" vs "illusory" probably isn't the best way to go. This stuff is hard, but it's more helpful if think about modes of existence being "ontologically objective vs ontologically subjective"; and phenomena as "observer-independent vs observer-relative".
English
‘Everything you see, taste, smell and touch isn’t real but an illusion. They are just electrical impulses.’ ?? Woeful BS from neuroscientist David Eagleman @BBCRadio4 Th Life Scientific who confuses perceptions with what they are perceptions of. Leave the philosophy to the pros
English
@neurobongo Indeed... although it does beg the question - to the next level of *what*, exactly?!
English
The actual, literal flex here is what takes it to the next level. 💪
Jürgen Schmidhuber@SchmidhuberAI
The most cited neural nets all build on our work: LSTM. ResNet (open-gated Highway Net). AlexNet & VGG (like our DanNet). GAN (an instance of our Artificial Curiosity). Linear Transformers (like our Fast Weight Programmers). people.idsia.ch/~juergen/most-…
English
@yudapearl What they sometimes don't realise (because the model-free phases of big projects can be so... errr... big) is that, often, the entire point of discovery science is to enable hypothesis generation, to *enable* model-*based* research in massively complex domains. 2/2
English
@yudapearl People sometimes get confused about this stuff. With the increase in funding of "big science" over the last 50+ years, there are research scientists who have been mostly exposed only to "discovery science", which is often essentially model-free. 1/2
English
I am retweeting because, again, the last line sounds punchy: "Science w/o conjectures would not be science." It aims to refute claims that model-free (ie, data-centric) scientists are somehow less biased, hence more successful than the model-guided scientists we all know.
Judea Pearl@yudapearl
@irinarish I prefer robots that communicate with us at the level of Newtonian physics to quantum robots who can't explain things: why pumps get hot when compressed. As to curve-fitting, see Greek-Babylonian rivalry here: ucla.in/2wj4pox Science w/o conjectures would not be science.
English
@WiringTheBrain I suspect the entire point *is* that it's fun (if you have a low bar for fun) and relatable. That is, the idea is to input into a commercial offering which aims to increase share price of 23andme e.g. 'We have determined that you have "got rhythm".'
English
GWAS like these are fun and all, but beyond confirming the heritability (and brain basis) of traits like this where does this research go?
Joe Pickrell@joe_pickrell
Genome-wide association study of 'having rhythm' biorxiv.org/content/10.110…
English
@WiringTheBrain @Abebab A more ambitious route…
1. Trademark “MegaCRISPR” age-reversal technology, and “InfiniStem” organ regeneration technology.
2. Found a biotech called “You Can’t Take It With You” and ask for $50bn to reserve a place in the queue to access the technology when it’s ready.
English
@Abebab On a scale from somewhat to very, how immoral is it to pretend to gullible investors that these approaches have any hope of working in order to land some of this sweet moolah? Asking for myself...
English
"Altos is luring university professors by offering sports-star salaries of $1 million a year or more, plus equity, as well as freedom from the hassle of applying for grants."
"reprogramming technology is risky & it’s a long way from a human therapy."
technologyreview.com/2021/09/04/103…
English
@WiringTheBrain Most of the thinking on representations in ML and AI is pretty unsophisticated, IMO (and that's before we get to the writing). This review is a bit old (2014), but it has the advantage of being intelligently written, at least. A good place to start, maybe: arxiv.org/pdf/1206.5538.…
English
I'm looking for some good reviews on the use of the concept of "representations" in computer science and artificial intelligence. Any recommendations?
English
@prokraustinator The depth of a NN (as well as other architectural features) obviously affects things like numbers of learnable parameters, inference multiply-add operations etc. These might be useful for some perspectives e.g. consideration of memory, computational efficiency, latency etc.
English
sciencedirect.com/science/articl…
NN folks, I'm working if/how it makes sense to talk about "the" depth of a neural network.
Per Cybenko and Hornik, all networks can be reduced to a 3-layer one, though it might be immensely fat. Distillation does work in practice though....
English
@cshperspectives @WiringTheBrain @DougCoupland The piece wasn't intended as satire (it's not very funny). It was, however, intended as clickbait - and, judging by the reaction across Twitter, it looks like it worked!
English
When @DougCoupland says Elon Musk “is measurably, scientifically, clinically and demonstrably the smartest person in any room anywhere”, I feel we need a reference ;-) theguardian.com/technology/202…
English
@neuralreckoning It's further complicated by the fact that, over time, as universities churn out more PhDs (see pic), expectations for what "good PhDs" look like change significantly.
But the short answer to your question: IMO, you can't reliably predict success in research by using exams. 2/2
English
@neuralreckoning It's complicated. Depends on which exams. Depends how you define "success" in research e.g. ability to: design and carry out experiments, write papers, operate independently, think of their own research problem areas; cope with failure and rejection etc. 1/2
English
Is there any evidence or reasoning that exam grades measure anything meaningful or relevant for research potential? I can only see problems with exam grades, but I still look at them when evaluating candidates, and let them affect my judgement.
English