I've read neuromancer a few times and it's one of those rare books that always holds up. It's a stream of words plugged directly into my imagination and I see that world in my own way, lingering around and filling in the spaces Gibson describes.
It's nice that people are putting a lot of effort into making a movie about it but for me watching would be like taking the tourist bus around a place I already know and love.
I'd feel something similar if they did The Mote In God's Eye, with the difference that I think curiosity would get the better of me there.
My high school kid reports that repeatedly on Chem tests a kid has his phone hidden under a calculator case and slides it out to AI the question, copies down then answer and shows it to the friends around him. And their marks are very high, higher than my kid who studies quite hard. I've said this will bite them in the future, but maybe if they get into uni ahead of my kid then I guess, maybe not.
In my kids words, the teacher is old and has no understanding that you can now take a picture with your phone and have a complete answer appear. I think he needs an update.
Our area just elected a new school trustee and I told him I have some thoughts. Ones now going to be retraining on cheating.
An excellent idea. There's a certain class of laptops that were never meant to be used on one's lap- at best merely portable e.g. Dell precision specced at 240w.
If the pad contact locks together for heat transfer it could also function as a security lock.
The dock could have a water cooling loop under your desk to dry your boots while you work.
This is exactly the case. On my Canadian high street the redevelopments are either empty or exclusively high cost businesses. I looked at the rent arrears notice posted on one failed shop and the rate would exclude anything not a money mill. They'll be no lutiers or second hand bookstores in the shiny towers on my street.
I would love to read about how insurance premiums are driving technology changes, in general. How insurance companies are tabulating risk from AI hacking. Risk of noncompliance probably brought the store operators to networked fridges, how is the network risk evaluated in general?
Insurance risk and premiums might be the main drag on the AI future.
An interesting question is why aren't there companies in the same vein as AMG was for Mercedes, that buy the product, modify and resell them? In this case, buy the factory LG and using expertise, lobotomize it and resell it people here? If this post shows the demand the where is the supply?
It doesn't seem that it's a DCMA limitation, yet at least. Would it fall under "right to repair"?
I think aircraft are so safe in this regard because, essentially, they are required to use a single international collision avoidance system that is well defined. And modern cars have AEB systems built by vendors aiming for commonality to meet worldwide rules.
But train traffic systems seem to be a hodgepodge of different systems, perhaps locally grown and driven by the local vagaries of politics and rulemaking.
What a surprising inversion of expectations regarding safety.
I've been in a near miss in a plane: the pilot in the other plane was clearly inexperienced, not following radio protocol, and we got within O(200m) of a head-on collision which certainly would have been fatal for everyone.
Thing is, the sky is big. You can have non-fatal near-misses that everyone learns from. In trains they literately railroad you into a head-on collision when things go wrong.
I believe that aviation is safe because of standardization, but there's also just a lot more space to almost mess up.
The newest signaling systems, ETCS and related sibling CTCS, are quite standardized and interoperable across international borders, so this is an area the EU especially has been making big strides on.
Cab-signalling train protection systems like LZB (and more modern alternatives like ETCS) are generally considered to be sufficient to ensure safe separation. The interlocking only allows setting a route if all its sections are detected vacant, conflicting routes are locked out, and flanks are protected. Meanwhile, the train protection system prevents trains from exceeding their movement authority. These systems can even largely monitor themselves and are designed to be fail-safe (often by putting all signals on danger and letting the operators figure it out).
But in those cases where the system does crash and reset, how can operations be started again? That's where you need well-specified fallback procedures. One such procedure might be that, in case axle counter sections to lose their counts (because of a crash of the interlocking or the axle count evaluator), causing affected sections to show as disturbed and mandating a manual reset, each section must be swept by a train running on sight (under a caution order in LZB or OS mode in ETCS).
So you detect them miles away. It's a real-time analysis problem: assign a latency equal to the longest time required to react and that's the time you have to detect an oncoming train.
Because rail is quite non line of sight and quite infrastructure defined pathing.
The automatic coordination system hit a fail condition; they neglected the fallback protocols that had rules for spacing determined to work reliably without the computer keeping real time supervision; they crashed.
This is in the way more like the Washington D.C. helicopter plane mid-air where some normalization of deviance got people to ignore automated protection systems.
If most involved had known the fallback protocols they would have refused to move/continue while the protocols were utterly neglected. At least with a robust safety culture active, ofc.
> This is in the way more like the Washington D.C. helicopter plane mid-air where some normalization of deviance got people to ignore automated protection systems.
I think that's not quite accurate. Yes, there was a normalization of deviance. But it wasn't primarily that automated protection systems were ignored.
First, the systems on board both aircraft could not have saved them. TCAS resolution advisories are inhibited below 900ft AGL and aural traffic advisories are inhibited below 400ft AGL. The aircraft collided below that altitude. The system wasn't ignored, it just doesn't do much at low altitude. The RJ crew got a TA about 20s prior to the crash, but that's not something that would normally require evasive action (they're trained to act immediately on an RA, not a TA). They were set up for landing. It's hard to find any fault with them.
The RJ wasn't equipped with ADS-B In, and the helo's ADS-B Out was misconfigured since installation. ADS-B In could have alerted both crews almost a minute prior to impact. But it wasn't available.
Second, if you're referring to the tower: Yes, the controller did get an automated conflict alert, but he did not ignore it. He asked PAT-25 whether they had the RJ in sight and told them to pass behind it. This response wasn't good enough, but that's mainly attributed to the controller's high workload and time pressure leading to omissions.
A proper traffic advisory should have included azimuth, distance, direction of travel, and type and altitude of the conflicting aircraft, e.g.:
PAT-25, traffic ten o'clock, one mile, north-west bound, a CRJ descending through four hundred, circling approach to runway 33.
A proper safety alert should also include an action to resolve the situation, e.g.:
Traffic alert, PAT-25, traffic ten o'clock, one mile, a CRJ on final for runway 33. Advise you [descend/turn] immediately.
Either one might have prevented the accident. But when ATC gets busy, the reality of those calls often looks quite different than how they're described in a textbook. Combining local and helicopter control with 12 aircraft on frequency is just too much. Of course, that leads to routine conflict alerts (5 separate events in the 18 minutes leading up to the crash). Sure, nuisance alerts lead to desensitization (and the report discusses that), but here the workload seems to have been the bigger underlying factor.
Of course, this is far from the only contributing factor. The normalization of deviance that the NTSB actually found are that visual separation had become the primary way of separating helicopters from aircraft, helicopter crews reporting "in sight" when they hadn't positively identified the traffic yet, the tower trying to "make it work" rather than forcibly slowing operations down, etc.
That in addition to putting a helicopter route as little as 75ft below the approach path, the helicopter's altimeter being 100ft off (within spec!), the helicopter crew's use of night vision gogggles severely limiting field of view, helicopters and fixed-wing aircraft being on separate frequencies severely limiting the RJ crew's SA.
And perhaps most egregiously, the FAA ignoring the countless warning signs (15 thousand close proximity events over 3 years, are you kidding me?) Failing to fix the helo route, failing to reduce arrival rate (the request from Potomac TRACON wasn't forwarded due to being "too political"), failing to improve the tower's conflict alert software...
People did not ignore automation, the regulators ignored lots of growing holes in the slices of Swiss cheese.
Would any teacher of, say, grade 4 students, reading this ask them to draw the same prompt and let us know the result? It would be so interesting to see how we differ from the machines, if at all.
That's both sad and hilarious thank you. I'm either expecting too much or people aren't being motivated (repeatedly trained by reward?) to turn in an earnest effort. Like the llms are trained that is.
I don't think they were prompted to make it a project, just a quick casual sketch. If a person doesn't use a stylus very often and doesn't draw or doodle ever then it could be surprisingly difficult for some.
This is the kind of thing that happens all the time in elementary schools. My mom taught elementary school for a while and her favorite thing was seeing all the students put their own spin on the art assignment(“prompt”, if you will).
Is every kid a Picasso? No, but there definitely is creative diversity among these kids who all have relatively similar shared backgrounds (neighborhood, social class, religion, etc)
I've pointed this out so many times. Even if you crank the temperature of the LLM up to the point where the oceans boil off into space, at the end of the day, generic in => generic out where prompting is concerned.
It's the reason that if you ask a random child to draw a house, you'll end up with the classic “house with a pitched roof”: a square plus a triangular roof, one door, two windows, and sometimes a chimney.
Re house - the reason why people (not just kids) will draw the same symbolically looking house is due to the question itself: they want you to recognize the house in what they've drawn. No one wants to fail, everyone wants to please, and it's the same with llms.
If instead you ask to draw some fancy interestingly looking house you'll get more variety - now the task is not to have the question giver recognize the house in their drawing, but to show how they imagine fancy or interesting house.
Well I wouldn't necessarily say that the LLM is trying to please you (that's a bit too anthropomorphic for my liking), I think it's more effectively that it's a "regression to the mean". Same reason why you can instantly recognize LLM prose.
Can you prompt around it? Naturally, but that's kind of my point. The more generic your prompt, the more generic the response.
In my experience, simply adding modifiers like "interesting, unusual, bizarre, imaginative" doesn't necessarily help either. Try asking an LLM for a "insert aforementioned qualifier adjective" fantasy short story in independent context/ conversations a few dozen times and you'll quickly see what I mean.
I wouldn't draw such general conclusions out of this, because it doesn't explain why LLM outputs are samish. Why shouldn't a generic prompt produce a mix of generic-but-notably-different outputs, as would be the case if you sampled humans? AFAIK, the answer is that it's RLHF that causes models to collapse to a specific style, because base models tend to score much better on creativity and output diversity evaluations than post-trained ones.
> Why shouldn't a generic prompt produce a mix of generic-but-notably-different outputs, as would be the case if you sampled humans?
Yeah, I hear what you’re saying. I guess it depends on how we define “notably different.” As I mentioned elsewhere, if you ask a random group of 12 people to draw a house on a piece of paper, you’ll end up with almost identical results. Sure, there will be some minor differences but for the most part, it’s going to be the same square with a triangular roof, two windows, one door, and maybe a chimney.
> AFAIK, the answer is that it's RLHF that causes models to collapse to a specific style, because base models tend to score much better on creativity and output diversity evaluations than post-trained ones.
Interesting. I'd love to see a paper on this. I wouldn't mind trying some back-of-the-envelope tests on a RHLF vs Base version of a locally hostable Qwen model too!
The first time I personally heard of RLHF ruining some model qualities was actually in the GPT-4 paper (march 2023); figure 8 there shows that the base model has near-perfect calibration on multiple-choice questions (!) and this stops being true for the RLHFed model: https://arxiv.org/abs/2303.08774v6
I think this is the same sort of problem - that the miscalibration of RLHFed models also explains their lack of diversity. In the training dataset, for a prompt like "you're a helpful assistant, yadda yadda, write a site", there's a large range of possible completions which are quite different and yet similarly probable, and if a base model is well-calibrated on that space, that lets you sample from that entire space and get many diverse websites. But an RL-trained model is miscalibrated and chooses a much smaller subspace of completions to assign unnaturally high logprobs to, so no matter how you sample, all the websites you get will look the same. It's not obvious to me why this happens; I wouldn't be surprised if, for example, a new optimization algorithm gets discovered that prevents RLHF from having such effects.
Sort of a tangent, but if you remember AI Dungeon when it first came out, it was based on a fine-tuned GPT-2 checkpoint which of course was pre-RLHF. While its adherence to user prompts wasn’t fantastic, it got a lot of praise for its diversity at the time.
I almost wonder if you could do something like this: have a non-RLHF model that gets the user’s request and just spits out a cloud of possibilities. It might not be the most tightly tuned, but then you could have a RLHF model take that output and use it as a sort of quasi-seed to increase diversity, while still retaining the RLHF advantages of producing an ultimate response that aligns with the user’s initial request.
That second "Mind the Gap" paper seems to be doing something very much like that, though their "conformative decoding" (CD) seems very simple: they make both the RLHFed and the base model produce a probability distribution for a given context, then apply top-p truncation to the RLHFed probabilities as one normally does, and then they look up what logprobs the base model assigned to those same tokens, and sample according to the linear combination of the two models' logprobs. This seems to work a moderate amount for them.
Looking at the cited-by I found a paper that evaluated a bunch of such methods by LLM-judged novelty, https://arxiv.org/abs/2606.12790 . CD is weirdly weak according to these results, they get better changes with other methods - but all of the others are less technical and basically amount to asking the model to try harder at being original. I wonder whether CD would work for a task like programming...
I could draw a better house. However it would be a lot more effort. (I'm not a good artist - a good artist could draw a better house in much less time than me - but I can draw that classic house much faster than the artist can draw the better house).
For most kids there isn't enough time to draw a better house - they want to go play. They also are well aware that there is only so much time in class and the better house wouldn't be done and in turn look even worse.
That would actually be quite interesting. But then there's something sad about the thought that the results, if shared online, would likely be used to train future models.
reply