Posts by Michael Wojcik 12317 publicly visible posts • joined 21 Dec 2007
More generally, terms like "smart" and "intelligent" aren't productive in sensible, informed discussions about ML and so-called "AI" systems, and even marginally more-specific terms like "sentient" (which is at best a niche attribute anyway) or "sapient" are unlikely to be of any real value. Reasonable ML researchers typically use terms such as "capabilities", which is general and too vague to be a metric but at least avoids some of the misleading connotations.
It's clear to most people with some expertise in the area that unidirectional transformer stacks on the scale of even the biggest LLMs publicly demonstrated so far lack many capabilities which humans display, and in fact that's only to be expected based on their architecture. (It's conceivable that some transformer-based architectures might spontaneously develop arbitrary cognitive capabilities if scaled up large enough and appropriately hyper-parameterized, Boltzmann-brain-fashion – conceivable, but by no means certain. But in any case we aren't near that yet.)
But even if you assign a pretty low paperclipping probability, or probability of AGI in general in a near timeframe, there's already the risks Hinton points to: LLMs are useful for bad actors of various stripes, because they can produce false evidence of decent quality1 and, with proper prompting, implement standard rhetorical techniques well enough to persuade many non-critical audiences. They work well as demagoguery amplifiers. They're also dangerous because they are reasonably good simulators of authoritative sources and careless (one might say venal) actors, notably Microsoft and Google, are pushing them as authoritative. That, too, imperils those who can't or won't attempt to think critically.
1In particular genres, of course. LLMs converge on a sort of median prose style, even modulo style prompting, which is pretty pedestrian; outside of that they're noticeably worse. Still-image generative models do very well with some prompts and less well with others. Video generation is ... a work in progress.
Yes, the pretense that San Francisco is somehow the most difficult place in the world for an AV to operate is prima facie absurd. It's certainly not the hardest place in California – there are plenty of rural areas where online maps will be inaccurate and signage negligible, for example, and where there will be hazards like fallen rock and washouts. It's probably not the most difficult urban area in California; I don't know of a worse one offhand, but there are a lot of them, so that seems statistically likely.
Try running a robo-taxi service in the region around the Mountain Fastness (hey, we have a commercial airport, so theoretically there's demand) and you'd find out that SF is a doddle.
In the UK, as you wrote, rural areas could definitely be a challenge. How are these AVs at negotiating fords? How do they do with blind intersections? Can I take one over Hardknott Pass? (When I were a lad, I spent a few days in Cumbria with my folks. We had a hire car, and my father and I were looking at a map and spotted the pass, so of course we had to drive it. Really very pleasant and not at all difficult, but not being able to see the road past the bonnet of the car was an interesting experience, and the only way you knew of oncoming traffic was the wave of sheep that preceded it. Possibly not accounted for in the AV's model.)
And while I've never driven elsewhere in Europe, my understanding from TV is that the UK is relatively easy.
Hell, even New England is a lot harder to navigate than San Francisco. SF is not that far from Phoenix (beloved of AV firms) in terms of difficulty.
Oh, do fuck off. Many people in the US are aware of working conditions and regulations in other countries, and their histories (which are far more complex than you suggest).
And the US didn't "abandon" slavery with the 14th Amendment; it nationalized it (and it's been privatizing it again in recent decades).
"naive" indeed.
And even if you can "get a job the next day", there's no guarantee that changing jobs won't adversely affect you. Maybe the new job isn't as close to your interests; maybe your coworkers or work conditions are less appealing. Maybe you'll have to relocate. Maybe the new employer turns out to be rubbish.
Of course, it might be better, too – but it's a gamble. Some people are risk-averse, and some others ought to be. And whatever happens, there's the cognitive load of taking on a new position in a new organization, and the attendant stress. The "you can always get a new job" (and its extended version "you can always quit and get a new job") philosophy does not realistically assess the costs of doing so, even for high-demand employees.
I suspect for many of the legislators this is the plan. Pass a virtue-signaling (for the reactionaries) law, knowing that in practice it will be ineffective so people won't be up in arms about its revenge effects. There will be the occasional wrist-slapping fine to show it's "working".
There's bread, and then there's circus. This is circus.
I attended and taught at a university on the fringe of the Bible Belt, and let me assure you there are a great many young people in the US who do, in all seriousness, consult religious websites. These were kids who ascribed to one flavor or another of Christianity, but I'm sure the same is true of many other faiths. Even teenagers are quite capable of being True Believers – even when such beliefs contradict others they hold. (Rationalizing cognitive dissonance is a skill most people master early.)
Once again, much of the Reg commentariat seem to have trouble grasping the idea that Not Everyone Is You.
I generally click the button1 and start typing, but about 30%-50% of the time it doesn't work – keyboard input is just ignored. I have to mouse away, click the desktop, and try again. I have no idea how they managed to screw that up so badly, but it happens on both my work and personal machines, so it's something to do with my configuration.
WIMP GUIs are ill-conceived and abysmally implemented in so many ways (modal dialogs, focus stealing, z-order issues, ...), but Windows always seems to go the extra mile.
1I used to use Ctrl-Esc rather than screwing about with non-keyboard controls, but the Win10 "start" button doesn't grab focus successfully when Microsoft's misnamed "X Mouse" feature (i.e. implicit focus) is enabled, so Ctrl-Esc is not reliable.
Why could they not be bothered to set up everything they need as subdomains of microsoft.com in the first place?
Because marketers love domain names the way flies love shit?
It's true that Microsoft has a huge problem with URLs, or really a number of problems. They've never really understood the concept, as anyone who uses Sharepoint (the worst information system ever invented by allegedly intelligent beings) knows. And like many corporations they've indulged in an addiction to acquiring idiotic domains, as you noted. But, really, using a word-TLD? The tackiest solution possible? What techie doesn't think word TLDs were an obnoxious cash grab by IANA and essentially useless?
It's more like an expert system.
A quick look at the NeMo Guardrails info in NVIDIA's technical blog post shows it's just a framework for writing rules in Colang, a data-flow language NVIDIA invented for this sort of thing. You basically tell it "if the input matches this pattern, do this thing", where "pattern" is a fairly expressive DSL. But there's no magic here, and frankly my initial guess is that if you have programmatic access to a NeMo-protected system (so you can drive a lot of queries) you'd be able to train an adversarial model pretty quickly to simulate its guardrails and then look for holes.
And, of course, it's up to the application developer to design good guardrails in the first place, and that's an area where very few developers are likely to show much expertise.
So this is likely to be used reactively: "Oh, someone just made our bot do a Bad Thing, so write another guardrail to prevent that specific prompt". No magic, no silver bullet.
(That said, a smart organization deploying an LLM-based service would do the adversarial modeling first, and use it GAN-style to create a set of guardrails. That's likely to be more successful than an economically-feasible human-authored set.)
When we moved and stopped paying for a cable television subscription a few years ago, it was around $250/month. That included Internet service, but now our combined Internet and streaming subscriptions is only around $130/month. (And it'd be less if it were just me here, since I could dispense with most of them, and would. Particularly those bastards at WBD.)
UK bans an American company from acquiring another American company. Yes I can see this being effective. Unless the 2 entities have a uk office, there’s not a jot the CMA can do about it except moan.
Always good to read a well-supported argument from an expert. Perhaps you know of one?
The no-fly list is a giant pile of crap, and "Elizabeth Holmes" doesn't seem like a particularly unusual name. Just have her surrender her passport; that should make leaving by commercial air travel sufficiently difficult. (If she can get a forged passport, she can get one under a different name, so the no-fly would still be useless.)
And that's why the industrialized nations have regulations regarding investment schemes and vehicles, and people and organizations which offer investment schemes and vehicles that fail to comply with those regulations are breaking the law. Oh, they say it's "novel" and "disruptive" and "we don't think your regulations apply"? Tell it to the judge.
There are a number of ransomware techniques that try to avoid basic detection methods, such as only encrypting certain types of files or encrypting files in a random order. That sort of approach might eventually trigger detection but possibly late enough that some damage would already be done. But of course earlier detection is always better.
There are also ransomware strains which hook the filesystem and decrypt on the fly to remain unnoticed until all files and recent backups are encrypted, and then spring the trap. So checking canaries through direct disk access is better – but of course gets more complicated with more-sophisticated storage systems. (Here checking backups routinely on an isolated, dedicated system – preferably automatically – can help detect "stealth" ransomware.)
It's the usual arms race. As with any arms race, though, even if you don't have great defenses against the state of the art, protecting yourself from the run of the mill is a good idea and gives you some breathing room for worrying about a really sophisticated attacker.
There's no way a microcontroller is running anything like a big ML model, so even by today's weak-sauce standards there's no "AI" here. There might be some processing of some sliding window of telemetry data by a canned model of some sort that's amenable to processing by a relatively low-power processor in an embedded system, such as an HMM. So, yeah, this is some arrant-nonsense marketing.
Anytime an American “group” buys anything, it’s about making money, fast. Fuck everything and anyone, history, values, customs etc. etc.
Not always.
US venture firms such as Golden Gate provided the capital that bought the Micro Focus unit away from Merant in 2001, unwinding the disastrous Intersolv merger of 1998. Merant was already in a tailspin in 2001, and Micro Focus thrived, paid back Golden Gate, went public again, and eventually bought the remains of Merant (by buying Serena in 2016, after Serena bought Merant in 2004).
That was a Good Thing for Micro Focus. I was there. It also allowed Micro Focus to buy Borland, part of Compuware, and the CORBA vendors, and to merge with Attachmate Group – and it kept those products alive (with a handful of exceptions they're all still around), which as we all know is often not the case after ISV M&As. And keeping the products available, supported, and updated 1 is what really matters to customers.
The merger with the software side of HP was definitely shakier, but again the products were kept alive.
And then OpenText bought MF this year, and shareholders made some money, and those products continue to be available and developed. Acquisition is always worrying but so far things seem encouraging.
Business is often ugly, it's true. But that doesn't mean taking a company private is always the road to ruin. Merant under Greenfield was a basket case and being taken private saved Micro Focus. Running a business takes money and when a public software firm is struggling to keep shareholders happy, an acquisition is probably going to happen regardless. So taking what looks like the best opportunity2 is not failure or capitulation.
(While I was on the inside for all of the Micro Focus saga, everything here is public information or opinion.)
1MF is sometimes accused of being a software retirement home, but nearly all the product lines – even ones with relatively small market share that are overshadowed by our own competing products – have regular updates with fixes, enhancements, and new features. Visual COBOL is the flagship COBOL product, but ACU and RM still have a regular release cycle. Visibroker is the flagship ORB, but Orbix gets updates too. ArcSight is the flagship but Sentinel is still getting refreshes. And so on.
2Of course this will be a gamble. It's a type of stopping problem, like the Secretary Problem.
avoid falling victim to "failures of imagination" – like the US did prior to the 9/11 terrorist attacks
This is one of the most annoying, if not particularly pernicious, myths surrounding 9/11. There was no "failure of imagination". The attacks were not only entirely foreseeable but a direct evolution of earlier aircraft hijackings and other attacks such as the 1993 WTC bombing. It was anticipated by the pilot episode of The Lone Gunmen (2001), which revolved around a plot to hijack an airliner and crash it into the WTC, FFS. It had not only been imagined two years earlier, but put in a widely-seen spinoff of a popular television show. That's about as successful as imagination gets.
The 9/11 attacks partially succeeded because US intelligence operations had moved too heavily away from HUMINT to SIGINT, and because practices were poor, particularly around inter-agency coordination. (Whether the DHS boondoggle has improved that is still an open question. Personally I think the DHS is a mistake, but it's very difficult to evaluate these things with any accuracy.)
One of the things, certainly, but it's not like they don't complain about things that have happened. DeSantis complains about children learning things from books, for example, and that's happened on occasion. Might even happen again. I mean, not in Florida, but in places where there are still sane people.
Due to the OpenText acquisition I don't have more-recent figures handy, but in the FY2021 Annual Report, the IM&G portfolio, which includes IDOL (the former Autonomy product) claimed around $390M in revenue. It's not broken down by product, though, and there are heavy hitters such as Vertica in that portfolio, so... you probably won't find out from published information.
It'll be interesting to see how it does under OpenText, where there's more of a relationship to their traditional product line, but also possibly more overlap with it.
I'll just note that HP's (Apotheker's) foolishness and incompetence does not, from a legal standpoint, excuse any fraud committed by officers of Autonomy. They can both be in the wrong.
So that shouldn't affect whether Lynch is convicted.
That said, I think Lynch should be tried in the UK (on the criminal charges; he already lost the civil case), and if convicted punished there. And I'd be in favor of dropping the whole thing except that seems rather unfair to Hussain, who after all is in prison already and was hardly solely responsible for whatever fraud did occur.
The real pity is that none of this blew back on Apotheker and the HP board that backed him. If I were an HP shareholder I'd be far more angry at them.
Sure. No one has ever abrogated a treaty. The US Federal government certainly never has.
Honestly, if Patel had said no to the extradition, what would the consequences have been? Some stern words from the US Department of State, some frowny press conferences from Justice, and in a week it would all be forgotten.
The UK government gave in when Patel approved the extradition. This was just the court saying there were no legal grounds to overturn it.
The real line of defense against the abuse of the extradition arrangement ought to be the Home Secretary. But the party that (some of) the voters put into power decided not to appoint one with any backbone, just one who would parrot the usual "MOAR SURVEILLANCE" bullshit.
Not that the US Federal government has been any better at defending civil rights of late, of course.
did HP's accountants even look at Autonomy's books? I don't know
All of this was covered in gruesome detail (I'd say "excruciating", but it was actually all pretty entertaining) in previous Reg coverage. HP chose KPMG to perform their due-diligence review of Autonomy's accounts; KPMG started and issued a preliminary report, but warned that they were unsatisfied with the information they had access to; HP's CFO raised objections to the deal but was overruled; Apotheker went ahead but later admitted he didn't even read the preliminary report.
In short, HP – under Apotheker's direction – just made a show of due diligence, rushed into the deal, and overpaid hugely.
Lynch and Hussain and others may well have committed fraud, based on the legal findings to date, but I agree that HP and Apotheker in particular bear most of the blame here. And also with the general sentiment that the extradition relationship between the US and UK has not been exercised in an equitable manner, or for the promotion of justice. I would not have minded the UK telling the US to take a hike in this case.
the agency "was taking the appropriate action to remove the company responsible for submitting the alert this morning"
"Remove the company"? What, with extreme prejudice? Is Florida about to bomb someone's headquarters? "We shall slay them all, burn their offices to the ground, and sow the parking lot with salt!"
"Together ... we have a real opportunity to" get bored with this and drop it in a few years. I mean, that's the whole point of Google, right?
It's just that in this case that would be a good thing.
(Though honestly that's true of most of the other stuff the Google toddler has tossed out of the playpen over the years.)
Actually, I'm beginning to fear it will survive. The determination of the Twitter addicts seems to know no bounds. Intelligent, rational (insofar as anyone is) people who ought to have been smart enough to leave it instead seem to be doubling down even as reports suggest it's getting objectively worse. I can foresee a much-reduced Twitter limping along for years on marginal revenue, with the remaining users congratulating themselves on having accurately predicted how Musk would "save" it.
Twitter could become the next Myspace. (Yeah, Myspace is still around.)
While transformer LLMs are not "a statistical inference machine" in any useful sense of that phrase, I'll agree they're not terribly impressive either. The big transformer LLMs are performing pretty much as I would have expected since "Attention is All You Need" was published. They're doing context continuation based on a transformer stack trained on a big corpus. What's to be impressed by?
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
