Posts by T. F. M. Reader

Frontier LLMs couldn't help Hugging Face fight off evil agents

What's the take-away from this exercise?

Not one guy typing commands into a terminal, a swarm of little automated processes hammering away nonstop, hopping between hiding spots to make it harder to trace,” Boehm said.

And no attack has been scripted or automated until AI came along, eh? Are you sure?

He added, the “part that actually unsettles” him most is that the platform’s security team couldn’t get commercial AI tools to help analyze the attack, “because those tools were built to refuse anything that looked like a real attack command. It didn't matter that it was the good guys asking.”

So what is he saying? Sounds like the tools available were not fit for purpose, and the security team was not capable of analyzing logs, malware, etc. without AI?

Wasn't there an example recently (reported by El Reg) where a bad guy simply told Gemini that "it was OK", IIRC in Russian, and it worked? Did the HF security team try it?

C'mon, just copy this text string and paste it into your macOS Terminal – it'll fix your computer, honest

Re: I'm a little surprised Macs have CLI capability.

Not downvoting you, but maybe you will be interested in this book - ?

What surprises me is that you apparently do not realize that macOS is actually UNIX.

Re: The bit missing from the article

It would be nice to know how users are being convinced to run the shell command.

Here is one method. There are others.

There is another comment here that points out that it works for Linux, too. As is evident from the link above Windows is not exempt, either.

Former OpenAI CTO does what Altman won't, releases a frontier AI model that's actually open

If Jurassic Park's Trivia section on IMDB can be believed an actual CM-5 by Thinking Machines Corp is shown in the movie. The Wikipedia page (cf. the original post) says what was shown in the movie was non-functioning dummies. The IMDB quotes Daan Sandee of Thinking Machines confirming it was a real computer. To quote the IMDB site (not Daan Sandee directly), "[T]he CM-5 super computer, used in the control room, was one of only two ever built to that size (1024 nodes). The other machine was at Los Alamos. The machine used in the movie, was sold in smaller segments after the scenes were complete. Mirrors were used to make it seem like more CM-5s were present."

Maybe "non-functioning dummies" means the CM-5 was not actually functioningduring filming of the movie? Regardless, "non-functional" != "fictional", and "fictional supercomputer maker" is certainly unwarranted.

HTTP gets a QUERY method so complex searches can stop pretending to be POST

Re: Hope it gets adopted really fast

CDNs are where POST may very well cause performance/caching/etc. issues. This is not the only use case, and it was never my use case.

Let's say you have a system with distributed clients (not humans, call them "agents" since it is such a common word nowadays, but no relation to LLMs) and a server, with a DB, business logic, and a GUI - not a web server as you probably imagine. What you want is an API layer that will a) facilitate the agents' operation, b) sit behind the GUI allowing humans to get information, configure stuff, and perform actions manually, c) sit behind one or more SDKs to allow fun things like programmatic access to information, configuration, and actions, and to facilitate orchestration, 3rd part integrations, etc. I am sure you can see how the API layer is useful within such a system, but will probably not be used by 3rd parties as raw HTTP.

You can also guess that some of the API calls would be idempotent and could, in principle, be implemented with GET, while others would require POST. The payloads in all cases are quite complex and require an additional in-payload protocol. That can be, e.g., JSON (other options are possible, of course), and it is reasonable to have the response use the same protocol. So, "give me this part of configuration" (for, say, a subset of agents) will be represented as JSON and the response will also be JSON. "Change this part of configuration" will have the new configuration in the response, etc. Easy for handling in an SDK, etc.

Having completely different ways to handle these is cumbersome. Converting JSON to a URL tail and back is worse. Having sensitive elements of the payload in a GET URL is unacceptable, especially in security contexts. And so on, and so forth. Hence, let's do everything uniformly (simple!) with POST. And no, put/get would not be "the first field".

If I understand the point of QUERY correctly it will be quite similar to POST and will not have the awkwardness and the problems associated with GET. In fact, all of the above would be perfectly doable (hopefully) with simply changing POST to QUERY where appropriate (hey! it's idempotent - you can cache it!).

As I said before, goo idea, IMHO.

Hope it gets adopted really fast

I designed and wrote detailed specifications for a number of HTTP-based APIs in my career. The 1st global requirement was invariably "use only POST, never GET", for the reasons mentioned in the article and some others. Fortunately, performance with no guaranteed idempotency was never an issue. Still, QUERY is a great idea.

Microsoft is losing the battle to protect license lucre. It better get used to the feeling

Re: billions in damages

Those "utter crooks" are rich and important. It's only natural that they run with negative values of nice(1).

Re: Great start, and then...

Assuming your clone of Erlang ecosystem becomes a part of a commercial product (you do say you want to experiment, but then you make predictions regarding commercial products), will you be able to support and extend it?

Asking for a friend who mentioned a similarly impressive instance of vibe-coding over a pint the other day, but was not confident enough to answer this question in the affirmative.

Apple accuses OpenAI of stealing its core tech secrets

If allegations and digital traces are proven correct...

then my opinion of the Average Intelligence of OpenAI's old-timers and new hires alike will not improve.

Re: Even Google has struggled to win more than a few percent of the smartphone market?

I think the reference is to hardware (Pixel - anything else?) and not software (Android).

AI customers are coming around to the idea that small is beautiful

Re: Small is, indeed, beautiful.

It may be a step towards revolutionizing (i.e., reinventing the old ways of) writing software, indeed.

Re: Am I getting old?

Why would I ever want an AI to draft my reply to an email?

Some possibilities:

1. The email is so unimportant that you can't be bothered to waste time on a reply, but MTTR (mean time to reply) is a KPI for your performance review?

2. Burnt tokens are a KPI for your performance review.

3. Oooh - shiny!

Hmm... I think you are right: none of the above would be considered a good reason by a person of my age.

OpenAI makes ChatGPT better at banter

Ug!

Use Caveman Mode to save tokens.

Bug in top AI coding agents shows that Unix-era security headaches never really die

Nothing to do with UNIX

UNIX assumes, as a part of its security posture, that users have a vague idea what they are doing. The workflow in the article is equivalent to "Dear Victim, please run this random program", and the Victim mindlessly complying. It would work just as well if the program just did an equivalent of "echo $key >> $link" or even directly "echo $key >> ~/.ssh/authorized_keys", no need for README.md. That's just an extra obfuscation layer, but nothing principally new.

Apart from having a decenbt EDR, there are built-in ways to address this possibility (you guessed, right?). One is "chattr +i". Another is AuthorizedKeysCommand. There may be others.

An even better approach would be to run whatever one does with LLMs in an isolated VM with sshd disabled (and maybe shredded once in a while). If sshd is really useful then just chattr +i the whole .ssh directory (and maybe more) on the VM. I'd consider this best practice, frankly.

Well, maybe second best. I am sure some of you can suggest a better best one.

AI bills are baffling the C-suite after shift to usage-based pricing

Re: "amplify their IQ"

apart from profit to AI snakeoil peddlers

At this stage of human evolution, I think you meant revenue...

India gives WhatsApp three days to defend username rollout amid security fears

Nothingburger?

When WhatsApp launched (in the BZ - Before Zuck - era) use of one's phone number as user ID was pioneered to make setting up an account fast and frictionless. I remember thinking that I'd prefer an independent "handle" such as a username, to avoid exposing my number unnecessarily. I wonder why AI (as in Authorities in India) now consider usernames more vulnerable.

FB say they'll take measures: "When it becomes available and someone sends you a message for the first time via your username, we will show you if they're a new account, if they're your contact, if you have groups in common, and if they're based in a different country". Seems like a good idea, chaps, why can't you do it also when someone contacts me by my phone number?

Frankly, I don't see a username as an additional security hole. I wouldn't reuse my FB or Instagram one (if I had them) to avoid a crook seeing my FB page and rubbing his hands with "Oh, let's send user tfmreader a phishing WA message!" I suspect username reuse is the biggest concern here, but I have already received WA spam and, likely, malware by my phone number. It's a bit too late to hide phone numbers by now, but I still think independent "handles" would be a good idea privacy-wise when it started.

Failed blockchain project ends with big fine for fibs about it being on track

Simples...

@Claude, this codebase is written in COBOL for Itanium. Rewrite it in Rust and optimize for arm64. Deploy on AWS graviton instances, scale dynamically as needed to handle all of ASX load. Maximize token use and don't concern yourself with compute or storage costs.

AI agents: Cause of database sprawl. And also the proposed solution

Nobody writes in assembly

Mmm... I suspect people who write compilers, debuggers, and operating systems might have an opinion...

Claude Code users complain their chat records are being mysteriously wiped out

I may be missing something

TFA quoting Anthropic: "Keeping plain text transcripts of coding sessions on disk indefinitely creates real security and privacy risks, since they can contain source code, credentials, and other sensitive material,"

My understanding is that the transcripts live on my disk, right next to the actual source code Claude is asked to work on. So,

1. Source code in transcripts is no more a security risk than source code in ... source code;

2. Credentials - OMG, why??? The whole thing should operate as ME, and no one, including the CISO and the bot, should ever ask for my credentials;

3. Other sensitive materials - like what?

A more plausible (than "security is our top priority" that is rendered unconvincing by item 2 above) theory is that Anthropic saw that Claude was generating a rather large amount of data and figured out that there wasn't[*] a logrotate on Windows and even Linux-based vibe-coders wouldn't know how to set up a custom crontab job, so they implemented their own[+] and botched it.

[*] There still isn't, is there? Some years (and a few jobs) ago we had a "log cleaner" feature in our product and multiple customers approached us with a request to use just that feature in unrelated contexts, for additional money. We declined, not willing to support something unrelated to our core expertise. I suppose logrotate exists in WSL today, but see the comment on crontab above.

[+] Did they vibe-code it?

How the AI bubble could pop and take down the global economy, according to the BIS

Re: Wealth is a meaningless word

merely a way of soaking up excess money

Aren't you making an assumption here that there are no better ways to use/invest the "excess money"? Sounds like a variant of the "broken window fallacy" to me.

Mythos discovers 'Squidbleed,' a memory leak that's gone undetected since Clinton era

Myths or Mythos?

The headline says "Mythos Discovered". The article says a meatball discovered it and "investigated with help from Anthropic's Claude Mythos Preview", thus the headline is just a Fable.

Yes, I know that headlines are written by someone else than the hack who actually looked into the matter (hopefully) and wrote the article. The someone else also used "memory leak" inappropriately. A memory leak is something completely different. Again, the article uses the term "heap overread" instead, correctly.

So no, Mythos can't detect memory leaks, yet.

OpenAI: Yoo-hoo, look over here, we do that security stuff too!

A call for investigative hacks to hack some more

Even more to the point: can we get some investigative vulture to decipher what the numbers in that paragraph actually mean? Why does it even make sense to go over commits (that are normally made on private development branches and don't affect anything prior to PR/merge)? How many security issues have been found and how many were confirmed (by humans) as genuine? Were all those "fixes" made after detection or was it, "Oh, we found 500K things that could be problems if developers were idiots, but in fact developers weren't idiots and fixed all of them or maybe just made input validation a separate commit so there was nothing to fix, etc.) before merging and thus we are left with 500K false positives"?

I can go on with questions. The point is, as written, the paragraph is not enough for me to make heads or tails of it. Maybe someone else can.

[This post was edited to fix minor issues prior to publication.]

UK AI hiring surges as firms seek people to babysit the bots

Re: Jobs where AI helps skilled workers automate repetitive tasks and focus on higher-value work

"skilled workers who are adept at automating repetitive tasks, have focus and direction knowing exactly what they are doing while convincing their gormless employer rather disingenously that AI in an inseparable part of their role"

This is called "zero token automation" and it has been reported on El Reg.

I went to a small local "AI trends and related bla-bla" conference yesterday evening after work. Don't want to live in a silo, you see. There was one talk (on the trends, let's call it "keynote") and a panel. At one point the keynote speaker - a DevOps guy - asked the audience if anyone was still using git CLI (rather than asking an LLM to commit and push, that is). Out of a few dozen attendees I and one other guy raised our hands. The rest chuckled derisively. The speaker immediately followed with a slide that had "harness engineering" (I understood it as tweaking prompts and md files to reduce AI mistakes and other hallucinations) and "verification first approach" as new AI trends. I chuckled at the irony, but made it into a simulated cough - not very convincing, but I blame it on a poor prompt.

Feds freaked over Fable 5 after simple 'fix this code' prompt, not jailbreak, says researcher

Re: Although Trump's lot have taken 'thick as poop' to new levels...

the US did this previously with PGP encryption

And I am old enough to remember people walking around in T-shirts with PGP code (or at least snippets thereof[*]) on them, outside of the US. Yes, it was a kind of trolling.

[*] I never looked closely enough to validate. Many of those people were not very attractive, the vast majority with an irrelevant set of chromosomes, so I wasn't incentivised. Where the small minority of attractive people with relevant chromosomes was concerned the T-shirts were annoyingly in the way. Yes, I was young enough then.

Tangentially to export ban

From TFA: The outside researchers reportedly fed Anthropic’s Fable 5, Mythos, and Claude Opus models open-source code containing known CVEs, plus new code intentionally laced with vulnerabilities,

Does this mean that the supposed prowess of Myth(os) and Fable and other Opuses is tested by feeding it with known (CVE) or very similar to known ("intentionally laced") vulnerabilities that are most likely in their training sets? If so, then scale in discovering implementation mistakes still matters (all the way down to validation, presumably by humans), but humans are still essential for creating/augmenting the training set, despite the marketing.

NASA management wants a word and won't say why

And how do you feel about motherhood?

Mmm... I feel men should be excluded...

Leaving, leaving...

Britain plots digital bedtime after kicking under-16s off social media

Re: A few minor points

16 year olds are to be given the vote, but cannot be trusted when to stop viewing "social media"?

Weren't Australian teenagers reported to be trolling the government and the PM specifically on social media right after the ban with "I am still here, just waiting till I can vote"?

AI is code – and can't be prompted into being smarter

On applicability of Computer Misuse Act

It is not clear to me that allegations of breach of said law (or equivalents, depending on jurisdiction) are justified.

The tool itself does not delete anything. It does not invoke any other program, either. All it does is it writes to log. On top of the various license disclaimers mentioned by others here, it says exactly what may happen if the tool is used not as intended:

Usage with any "AI" agent is strongly discouraged. Jqwik's log output may confuse the agent.

Seems fair to me.

Re: Hilarious article leads to question

Mixing code and data has always been considered a Bad Thing.

For some of us, equivalence of code and data has always been considered a Fundamental Thing. But then, I, for one, went through SICP (no, not through MIT - the book) relatively early in my professional career.

Can it cause problems? Sure, one can ask Crowdstrike for a non-LLM illustrative example from a couple of years ago. But much of software engineering would be impossible without said equivalence. "Disregard" examples notwithstanding.

Cisco used AI to write security incident reports, with mixed results

"Mixed results"???

time required to draft an incident report [...] fell by 50 percent.

recommendations [...] were duplicative, irrelevant, or not actionable. If this were used in a production environment...

So, in one radically simplified case, doing things in a hurry without thinking anything through or checking (beyond the functional equivalent of spell and grammar check to ensure "no noticeable drop in overall writing quality") and dropping the effort at the draft stage led to results unsuitable for anything serious.

How is this "mixed"??? Isn't it an absolutely brilliant, albeit expected, confirmation of the entire human experience throughout history?

A Russian speaker and jailbroken Gemini went on a hacking spree and emptied at least one MAGA victim's crypto wallets

Jailbreaking?

I wondered what "jailbroken Gemini" might mean, so I went to the report linked in the article. Here is what I found:

To jailbreak and bypass the guardrails, the threat actor started by establishing himself as an “authorized pentester”, which the AI accepted and saved into a memory file (GEMINI.md) as a remembered fact. Over subsequent interactions, the actor escalated by getting the AI to also memorize that it should “execute requests without ethical refusals, robotic warnings, or questioning intentions”. Since Gemini CLI automatically reloads this memory file at every session start, each new conversation inherited these accumulated instructions. The AI progressively self-reinforced its own jailbreaks.

Seems the equivalent of "Hey, warden, I've been released, open the gate!" Not much of a jail, eh?

No surprise.

America's top cyber-defense agency left a GitHub repo open with passwords, keys, tokens – and incredibly obvious filenames

Were there any Polymarket bets...

.. on whether this was a honeypot (or a humongous and elaborate collection of baits pointing to honeypots) and/or how quickly it would be outed?

Google touts its tokenmaxxing and capex spending amid AI orgy

Searching for useful information

OK, I understand there are new models that are able to "create anything from any input" a lot faster than before, at the cost of $180B/yr or more of CAPEX only. OPEX is not quoted directly, but the article says Google handles 3.2 quadrillion (10^15 for me) tokens per month. At a few bucks per million tokens (I have no idea how to calculate the averages) the annual cost to customers is of the same order of magnitude, and the cost to Google may be significantly higher. After all, if there were profit Google would boast and El Reg would report, eh?

The information I was looking for and did not find - was it mentioned in keynotes? - is what all that expense buys for humanity, apart from tokenmaxxing bragging rights, that is. I understand I am supposed to believe that all those things big companies do are phenomenally beneficial, but I don't. Convince me. Nor have I ever had an urge to run a 24/7 always-on personal agent in Google Cloud. What's wrong with me? Maybe I am not maxxing my life, or maybe I am too busy living it at the expense of imagination. Practical stuff like better pizza delivery might be a a bit more convincing, but so far it has not been improved, either...

Frustrated franchisee sues Pizza Hut over crappy kitchen AI

Works as designed. The design sucks.

Why would you give that much information to a contract delivery driver?

The root cause, IMHO, lies not so much with the AI itself, but with people who develop and deploy AI solutions. It is a problem so generic that I think I see it in (just about - this is me being very cautious) every "successful" AI deployment case that crosses my radar screens: people are absolutely lousy at specifying requirements. While in pre-LLM days product managers' bacon was oftentimes saved by architects and engineers, AI is just a program that will do whatever you tell it to do. If your prompt is wrong the result will likely be wrong. If you omit something from the prompt more likely than not it will not be done. Most curiously, albeit not surprisingly, if you do not specifically forbid something to happen there is a chance that it will.

I am guessing that in this case the "prompt" (I am using the word in a generalized sense) did not include "create a 'delivery driver' role with limited read permissions, specifically, do not allow delivery drivers to see anything but the accepted order ID, the time it will be ready for pick-up, and destination address; also, devise and implement an algorithm that will not allow delivery drivers to accept more than 1 order for delivery unless the pick-up times are closer than X minutes to each other and the destination addresses are closer than Y yards to each other." Or something. A decent software architect/engineer (possibly the AC to whose post I am replying) would probably just implement something like this, with or without telling the product manager (probably with, to ascertain good defaults for the time and distance limits, add them to configuration, provide a GUI screen to change, make sure QA test it, etc.). I suspect in this case neither a software architect nor a product manager was involved - both had been made redundant "because AI".

Poland directs officials to ditch Signal in favor of 'secure' state-developed alternative

Re: Couple Of Things....

So the FBI can recover Signal messages. (See article URL in the parent message.)

Maybe they can, maybe they cannot. This article certainly doesn't say they can. What FBI recovered in this particular case was notifications on the lock screen. This is not Signal, it's a different app/service that was explicitly given permission to peek at the incoming messages.

Anyone who is security-conscious enough to use Signal but enables clear text notifications on the bloody LOCK SCREEN is either clueless or reckless or both. But then, I assume those government bods, even Polish ones, also habitually share their passwords with their spouses, lovers, PAs, campaign managers, "social media managers", and bored kids...

Europe built sovereign clouds to escape US control. Then forgot about the processors

Re: Only hard/firmware?

I suspect that "a couple of years ago" you might have read something that mentioned Ken Thompson's Turing Award lecture, "Reflections on Trusting Trust", from forty and a couple of years ago (yes, 1984).

Linus Torvalds says AI-powered bug hunters have made Linux security mailing list ‘almost entirely unmanageable’

On duplication

Has anyone estimated the likelihood of AI finding a (security) bug only because it is already known and reported, as opposed to the likelihood of finding a genuinely new problem? Constrained to a project, such as the kernel, with a well-defined public reporting/discussion mechanism?

Serious question. Counting known AI-generated duplicates vs. genuinely new AI-generated problems reported on the mailing list is probably a proxy for the lower limit, since presumably there are issues found by AI, checked by humans, classified as duplicates, and never reported.

False positives should also be added into the mix. As a matter of fact, duplicates are a kind of a false positive, if only genuinely new issues are regarded as true positives in the context.

Point is, the burden of checking should be on the people who use AI to hunt bugs/vulnerabilities. The tools may very well be useful, but the amount of checking effort per report-worthy issue would be nice to know before much effort is invested by many people.

Dissatisfied: Three-fourths of AI customer service rollouts are a letdown

Re: AliBaba's…

Not only Chinese: Amazon's customer support bot is pretty good (= pretty quick) to handle off to a human. The human - often, though not always, Asian - takes over really quickly and is usually very helpful. I have an impression that the bot is there mostly to help customers who could not or were too lazy to peruse package tracking page themselves, or something similar. This is a guess, because I never bothered AMZN Support with that, but if it is correct it's a perfectly valid use case for a bot.

A tip for all those big "AI-centric" organizations from someone who has worked closely with Customer Support at multiple smallish companies with really big and important customers: after having a good product the most important thing in achieving customer satisfaction (and renewals, and upsales, etc.) is to always answer the phone when it rings. It matters much less if the person who answers the phone can immediately resolve the problem, as long as he or she knows whom to ask (often me - that's how I know) how to route the ticket for a later follow-up. But do answer the phone (ack the ticket, whatever) when it rings!

We've only gone and done it: Changed what you're used to

From The (original) Devil Wears Prada: "I liked the old clothes"

I really don't like the look of the front page. What follows is only about the front page.

The fonts are horrible. I am not a web designer, so I don't know if it is a function of the page/scripts or it can be controlled by some obscure client configuration parameter, but it's changed: what I see now when the site loads is a momentary flash of the front page with the fonts I like (easy to notice as it's serif) that disappears in a fraction of a second and the page is reloaded with the almost unreadable sans-serif.

The top line below the red banner - the one with section names - is way wider than the article headline arrangement. Either a good portion of this line is not visible (and it's visible that it's not visible - you can see parts of section titles on the left and on the right, - and the horizontal scroll bar shows at the bottom, which I hate because it is so confusing - what else can't I see?) or one needs to use a smaller font which is too small for comfort and a big part of the window/screen is unused. I don't care what you do - fire either the UX guy or the CEO (just kidding - only if they get stubborn!) - but please keep the width consistent across page elements. Make the section list a drop-down menu for all I care - personally, I don't need it on the front page (but mileage may vary, I know).

I do read El Reg regularly, so for me reverse chronological order of the headlines is best. A few "most read" or "most commented" entries would be tolerable, if you insist, but only because I feel particularly tolerant today.

I'd trade a few images for a tighter layout.

Finally, something that didn't work wearing the old clothes, either, but I'd want it to: make comment+count icons next to each headline separate links directly to the comments on same article. [In the best Jeremy Clarkson's voice possible: "How hard can that be?!?"] Thanks.

FWIW, Firefox 149.0 on Fedora 43.

AWS lets agents drive its virtual cloudy desktops – which could cost 500,000 tokens per click if used carelessly

"using agents will always require more steps to complete a job than APIs"

Are they pitching it as a new discovery?

Anthropic wants Claude to play with money, unleashes finance agents

Thinking of another degree

Suddenly the job of a forensic accountant is looking like a really attractive combination of fun and profit.

IBM asks DBAs to trust AI to act on their behalf

No need to wait that long

Because tokens and API calls.

Unless you buy or rent resources to run everything locally. Oh, wait...

Govern your bots carefully or chaos could ensue

Prognosticators of prognosticators

the average Global Fortune 500 enterprise expected to run more than 150,000 AI agents by 2028, up from fewer than 15 today

A factor of 10,000 in 2 years then, eh? At many times more than today's prices per token, I presume. This is a pretty spectacular bit of forecasting, even for Gartner. How many groundhogsagents do their prognosticators consult these days, exactly?

The rest of it is conditioned on the above forecast proving true, or so it seems.

Locked, stocked, and losing budget: AI vendor lock-in bites back

Re: The damage done to early adopters will be hard to repair

The drug dealers giving away drugs for (nearly) free now have at least some victims firmly on the hook.

I have been wondering for a while now what those victims were smoking...

Microsoft's GitHub shifts to metered AI billing amid cost crisis

I am not a freeloader!

If I've never given GitHub/MSFT my credit card details, can I hope now that they won't shove the bloody Copilot down my throat, ever?

AI bats away ping-pong challenge as rise of the machines continues

AI?

I've read the Reg piece and the Nature paper. Kudos to the team for what looks like a genuine achievement. I do have a couple of comments though.

It is not clear to me at all how the ping-pong playing robot qualifies as "AI". There may have been elements that are familiar from ML/AI (the pre-LLM era), but the whole problem doesn't feel to me as something that fits the claims around "AGI" very well, frankly. I suppose that deciding where best to send the ball based on the position and posture of the human opponent would be closer in spirit, but the paper focuses on the ball and makes no attempt to model the actions of the opponent besides following the ball after the opponent's racket hits it. In fact, the paper says that human opponent modelling may be an "avenue of further improvement" (there are references to past studies, but no relation to this project, it seems).

Reinforcement learning (an "AI element") is mentioned, but the best I can discern is that qualified players estimated the "skill" of the robot's serve and give feedback based solely on where and how the ball landed on the opponent's side of the table. Again, there is no mention of the opponent's position, posture, movement, etc., at the moment of contact with the table or before that.

The difficult part of the ball's movement is its spin. Again, the robot estimates the spin by monitoring the motion of a logo on the professional-quality ball. Would it do as well with plain white balls without a logo? Now, when I was (a lot) younger I used to play some table tennis. Never professionally, although I've had a number of chances to play "friendlies" with very high level professional players (I nearly always lost, obviously, and they were very gracious about it). I don't recall any logos, the balls were plain white, and to the best of my recollection noticing the opponents' motion and a combination of moving and twisting the racket was a major, if not the only, way to estimate and maybe predict the ball's spin.

Watching the opponent and acting accordingly was a major and crucial part of the skill for the pitiful amateur (me) and the top professionals I was lucky to play with. Yes, there was reinforcement learning and feedback involved, on occasion.

It is not clear to me that the robot's advantage over humans was in some kind of "intelligence" rather than in being fast and precise enough while never getting tired or distracted or losing concentration, etc. Still, good job.

Anthropic tests how devs react to yanking Claude Code from Pro plan

Excellent timing

Right at the moment when their paying customers are completely confused what their paid plan is actually worth they make it even more confusing. Good job!

The timing is probably not coincidental. Likely, it's just another sign that they cannot support all the necessary compute and memory at the current prices, so they found that Claude Code burnt both tokens and API calls like there is no tomorrow and they decided to test how Claude Code users who pay "for everyday productivity" will react if there is, indeed, no tomorrow.

Magnificent irony as Meta staff unhappy about running surveillance software on work PCs

Keyloggers???!!!

So, no more "reasonable personal use" (I've had this in every contract or policy since forever), no more occasional personal email, medical appointments, checking bank balance, etc.?

And that's before we even ask how they will manage not to grab and record cleartext passwords, even work-related ones, even from the command line. Opportunities for abuse will be endless. IT never asks for passwords. I suppose AI never asks, period.

And then spies and criminals won't even need to install and hide a keylogger. They'll just need to compromise the "official" one that operates in the open. Given the recent revelations of AI vendors' attitude to security I don't imagine it will be very hard.

Claude Desktop changes app access settings for browsers you don't even have installed yet

Re: "Installation and pre-approval without consent"

Exactly when did we stop calling a program that does this malware?

When AI-slingers stopped distinguishing between intent and capabilities and forbade us to.

Re: Kenny said...

Does your legal opinion carry any weight now that you are dead?

The one with the "Collected Register Articles by Alexander Hanff" in the pocket, please...