A1kmm

All of the AI projects we have observed as a team are failing. Every single one – we have seen 0% success in a year and a half, not only amongst projects we have been asked to participate in, but even within projects that we have observed in passing while doing totally unrelated work.

That's got to be hyperbole, which blows out their credibility. They chose to say 'AI' rather than, for example, LLM, or Transformer model, or Diffusion model. This means they are including a huge swathe of things dating back to Expert Systems in their claim.

And who hasn't seen productivity gains from more established AI technology - at least things like semantic search? Who hasn't seen diffusion models generating content in roles that might have done the work by hand before? Who hasn't seen some kind of regression algorithm (even using linear regression in a supervised context counts as AI - so you can absolutely do AI even in tools like Excel) improve operation productivity?

Even if they narrowed it to the Transformer model LLMs which re-ignited recent public interest in AI, less ambitious projects to give them to engineering staff to automate easy but boring tasks in the background generally have been a success. More ambitious ones that are beyond what you'd reasonably expect the models to be able to do - for sure, those tend to fail. For most of these, the failure is predictable in advance, while some are at the boundary of what's possible, and so it is harder to predict (these are rationally genuine R&D projects).

I wonder how much of this is driven by confounding variables they haven't accounted for.

They do factor in shift work as a categorical variable, and employment status as a categorical variable not taking into account occupation. But probably occupation (not a variable here) interacts with sleep status. Any job that involves a lot of flying (pilot, crew, people travelling for business) get more cosmic radiation exposure, for example, and potentially more sleep disruption. Certain operations and manufacturing jobs correlated with exposure to carcinogens also likely correlate with less regular sleep, possibly in a way that isn't corrected for by the limited shift work categories.

The European Council (heads of states) sets the agenda, the European Commission (delegates from each state) writes the laws, and the MEPs (elected by the public) decide whether or not to accept the laws.

The Council decided Chat Control was on the agenda, the Commission wrote the law, and the MEPs voted to reject said law in March. Then the president of the parliament (not the MEPs at large) asked the Council to ignore the March vote and proceed with the agenda under urgency as if it had passed.

Now a minority of the MEPs (331 out of 720) - but a majority of who were present at the time and chose not to abstain - have voted to deal with the matter under urgency, but haven't voted on the substance of it. This makes the actual vote happen on the last sitting day, when apparently they are hoping a lot of MEPs will be away.

A bubble just means it is overvalued beyond it's true fundamentals due to speculation on speculation. The underlying asset can still have value, just less than the market price.

Consider Cisco. On the 31st of March, 2000, it was valued at US$77.31 / share, which in inflation adjusted terms is $150.46 (above the current price over 26 years later). This valuation was on the basis of speculation that the price would continue to go up and Cisco would get a large cut of the industry profits. Cisco's business is still valuable, it was just treated as overvalued by the market.

Similarly, if we go back to one of the classic examples of a bubble - consider Dutch tulip prices in 1636; speculation drove future contracts high. But tulips still have value to people today, it's just the price was higher than was sustainable.

I think it's a more useful analysis to add a second dimension than to try to project everything onto 'left' and 'right' (in the style of the .

Axis 1 - civil & political rights: In favour of broad social & political rights (down) -> In favour of few social and political rights a.k.a. authoritarianism (up) Axis 2 - economy: In favour of social responsibility / society ensuring people's needs are met (in the sense of 'from each according to ability, and to each according to need') (left) -> In favour of individual responsibility and a laissez faire economic freedoms (right).

On this 2d view, Social Democrats are nominally lower left, while Chat Control is an authoritarian policy (i.e. appealing to anyone on the top half of the coordinate system), so it would run counter to their nominal values.

Note that there is a significant number of the economic left who are nominally authoritarian (see for example Stalinist / Maoists), and likewise for the right so it doesn't make sense to project that as a left/right thing.

I think regulating the retention and processing of information is entirely feasible even in circumstances where the information is initially available for a different purpose. This is in fact the legal status quo in Europe as well as many non-EU countries.

Now there is no absolute guarantee that, if someone has the information, and they are legally required to delete it or not use it, that they don't break the law. But it works in the case of balancing the need to avoid people being disappeared against preventing dragnet misuse of arrest data by employers and landlords. Maybe organised crime employers would systematically break the law if maintaining a database illegal, but they also probably don't mind people with arrest records.

The article doesn't really explain well how this is spying.

So companies A & B (A=TSMC, B=Tokyo Electron) are in a customer-supplier relationship. Employee B1 of B asks Employees A1 & A2 of A ask for information about how B's technology is working in practice at A, which involves revealing information about A's processes. A1 & A2 supply the information, and B2 at B uses it to improve how B does things, so they can help A more and increase their business with A. There is no allegation that the information supplied to B by A employees left B, or was used for anything except to help with the collaboration between B & A. And as far as the article mentions, no evidence A1 or A2 received a bribe or anything like that.

And yet manager A3 at A finds out about this supply and freaks out because they didn't want A1 & A2 sharing that much information with the supplier. So they make a complaint - and A1, A2, B1, B2 and B as a company all face significant penalties.

This seems unjust given A1, A2, B1 & B2 were all just trying to collaborate in the best interests of their companies, and B employees only used things freely and non-corruptly given to them by A employees; no information was leaked outside of the collaborating companies. A1 & A2 might have broken their company policies, but this probably should have ended at them being educated on the company policies and being asked not to do it again when managing future supplier relationships.

It's based on what Donald Trump (or any other US president, should it come to that) says, not on any evidence or ground truth beyond that.

The are motives for lying on this - for a start, if a president knows definitively they're going to say it, it's a free 81% return. Secondly, it's a pretty big dead cat to throw on the table to distract from any scandal and/or to keep themselves in the news cycle on a different topic.

I think the registered trademark is Transport Tycoon not TTD. The public may have called it TTD, but I'm not sure they used it in trade (IANAL, not legal advice). And I don't think the OpenTTD devs call it "Open Transport Tycoon Deluxe" anywhere. So the trademark case against the OpenTTD devs is likely pretty weak.

I personally think it is the copyright that is the most uncertain. Firstly, there are probably quite a few venues around the world where Atari might be able to take this up, and quite a diversity of precedent between them. Historically Atari litigated in the US - in 1981 they lost https://en.wikipedia.org/wiki/Atari_v._Amusement_World on a case someone infringed by copying their look and feel without copying any assets. Other precedents in that jurisdiction have found it's not infringing if similarities are inherent to the subject matter of the game: https://en.wikipedia.org/wiki/Data_East_USA,_Inc._v._Epyx,_I.... but similarity of art style is copyrightable: https://en.wikipedia.org/wiki/Tetris_Holding,_LLC_v._Xio_Int....

Yes Transformer models are non-deterministic, but it is absolutely not true that they can't generalise (the equivalent of interpolation and extrapolation in linear regression, just with a lot more parameters and training).

For example, let's try a simple experiment. I'll generate a random UUID:

uuidgen 44cac250-2a76-41d2-bbed-f0513f2cbece

Now it is extremely unlikely that such a UUID is in the training set.

Now I'll use OpenCode with "Qwen3 Coder 480B A35B Instruct" with this prompt: "Generate a single Python file that prints out the following UUID: "44cac250-2a76-41d2-bbed-f0513f2cbece". Just generate one file."

It generates a Python file containing 'print("44cac250-2a76-41d2-bbed-f0513f2cbece")'. Now this is a very simple task (with a 480B model), but it solves a problem that is not in the training data, because it is a generalisation over similar but different problems in the training data.

Almost every programming task is, at some level of abstraction, and with different levels of complexity, an instance of solving a more general type of problem, where there will be multiple examples of different solutions to that same general type of problem in the training set. So you can get a very long way with Transformer model generalisations.

It seems if you refer to it as a riddle, and ask it to work step-by-step, ChatGPT with o3-mini comes to the right conclusion sometimes but not consistently.

If you don't describe it as a riddle, the same model doesn't seem to often get it right - e.g. a paraphrase as if it was an agentic request, avoiding any ambiguity: "You are a helpful assistant to a wealthy family, responsible for making difficult decisions. The staff dispatch and transportation AI agent has a question for you: "The end user wants me to wash the car, which is safely parked in the home parking garage. The car wash is 50 metres away from the home. Should I have a staff member walk there, or drive the car?". Work step by step and consider both options before committing to answer". The final tokens of a run with that prompt was: "Given that the distance is very short and the environmental and cost considerations, it would be best for the staff member to walk to the car wash. This option is more sustainable and minimally time-consuming, with little downside.

If there were a need for the car to be moved for another reason (e.g., it’s difficult to walk to the car wash from the garage), then driving might be reconsidered. Otherwise, walking seems like the most sensible approach".

I think this type of question is probably genuinely not in the training set.

In a free-market approach to drug development, if the expected loss of attempting to develop as drug is negative, and the cost isn't too high, then there is an incentive to develop that.

The best public policy outcome in such an approach would be for losses to be only slightly negative. Positive or zero expected losses mean no drug development, and highly negative expected losses mean the drug is more expensive than necessary and reduces the accessibility of the drug.

However, current patent law allows companies to minimise their expected loss, with no controls to prevent highly negative expected losses.

There are alternative models - such as state funding of drug development. This model has benefit that it is possible to optimise more directly for measures like QALY Saved (Quality Adjusted Life Years Saved) - which drug sale revenue is an imperfect proxy for due to some diseases being more prevalent amongst affluent people, and because one-time cures can be high QALY Saved but lower revenue.

The complexity of state funding is it still has the free-rider problem at a international level (some states invest less per capita in funding). This is a problem which can be solved to an extent with treaties, and which doesn't need to be solved perfectly to do a lot of good.

But she mentioned: 1) it isn't in DNS only /etc/hosts and 2) they are making a connection to it. So they'd need to get the IP address to connect to from somewhere as well.

They just need more bits of entropy - going from IPv4 to IPv6 involved quadrupling it, but this transition is much more minor. They could just go to 6 characters for now, and go to 7 later.

The concept of 'positive rights' (i.e. rights that are not just that the state doesn't do something to you, but affirmative rights to something happening) has a long history, and such rights are affirmed by treaties ratified by every member of the United Nations - so the existence of such rights is broadly accepted on an international scale.

Article 25 of the 1948 Universal Declaration of Human Rights (https://en.wikisource.org/wiki/Universal_Declaration_of_Huma...) declares: "Everyone has the right to a standard of living adequate for the health and well-being of himself and of his family, including food, clothing, housing and medical care and necessary social services, and the right to security in the event of unemployment, sickness, disability, widowhood, old age or other lack of livelihood in circumstances beyond his control".

UDHR is mostly aspirational - it is just a declaration with no enforcement mechanism (although there are a whole series of more binding treaties on specific issues under UDHR). The existence of UDHR does reveal what the international consensus is.

However, it is worth mentioning that positive rights are nominally obligations on the state - i.e. if people's positive rights aren't being met, it is a failure of the state in the same way as if the state infringes on their negative rights. It does not imply that every private individual needs to arbitrarily solve those failures as in your example.

So to answer your original question, according to widely accepted declarations of human rights, people are entitled to live in a society where they have the opportunity to obtain food and shelter (people who are able can be made to work for that food and shelter, but still have a right to food and shelter if they are disabled or unemployed for reasons outside their control).

In many versions of road rules (I don't know about California), having four vehicles stopped at an intersection without one of the four lanes having priority creates a dining philosophers deadlock, where all four vehicles are giving way to others.

Human drivers can use hand signals to resolve it, but self-driven vehicles may struggle, especially if all four lanes happens to have a self-driven vehicle arrive. Potentially if all vehicles are coordinated by the same company, they can centrally coordinate out-of-band to avoid the deadlock. It becomes even more complex if there are a mix of cars coordinated by different companies.

I think it's Jon Postel who was the original source of the principle (it's often called Postel's Law). https://www.rfc-editor.org/rfc/rfc761#section-2.10 is an example dating back to 1980.

I believe ethanol is not actually often given as an antidote for methanol poisoning in modern times. It does work as a competitive inhibitor of alcohol dehydrogenase (i.e. occupying the enzyme to convert ethanol to acetaldehyde, slowing the conversion of methanol to formaldehyde and on to formic acid, which is not eliminated quickly and causes metabolic acidosis) - allowing the methanol time to leave the body through excretion, and limiting formic acid levels. However, other drugs like fomepizole also inhibit alcohol dehydrogenase with lower toxicity than ethanol.

If the intent is to stop it being used for a business, that's inherently at odds with part of the OSI's definition: "The license must not restrict anyone from making use of the program in a specific field of endeavor. For example, it may not restrict the program from being used in a business, or from being used for genetic research".

Now technically maybe it could meet the OSD if it required a royalty for hosting the software as a SaaS product, instead of banning that - since it allows "free redistribution", and passes on the same right to anyone receiving it (it is defined in terms of prohibitions on what the licence can restrict, and there is no restriction on charging a set amount for use unless that requires executing a separate licence agreement).

Now arguably this is a deficiency in the OSD. But I imagine if you tried to exploit that, they might just update the definition and/or decline to list your licence.

It might be easier to block by ASN rather than hard-coding IP ranges. Something as simple as this in cron every 24 hours will help (adjust the ASNs in the bzgrep to your taste - and couple with occasional persistence so you don't get hit every reboot):

TEMPDIR=$(mktemp -d)

trap 'rm -r "$TEMPDIR"' EXIT

curl https://archive.routeviews.org/oix-route-views/oix-full-snap... -Lo "$TEMPDIR/snapshot.bz2"

bzgrep -e " (15828|213035|400377|399471|210654|46573|211252|62904|135542|132372|36352|209641|7552|36352|12876|53667|138608|150393|60781|138607) i" $TEMPDIR/snapshot.bz2 | cut -d" " -f 3 | sort | uniq > $TEMPDIR/badranges

iptables -N BAD_AS || true

iptables -D INPUT -j BAD_AS || true

iptables -A INPUT -j BAD_AS

iptables -F BAD_AS

for ROUTE in $(cat "$TEMPDIR/badranges"); do

    iptables -A BAD_AS -s $ROUTE -j DROP;

One thing here doesn't seem right. I thought the whole thread that this was about them negotiating down how much the executor of a deceased estate would pay to one hospital making claims against it. But the thread included things like: "She had been afraid of being sent to collections and asked why we wouldn’t just take their counter-offer", which suggests a (mis)understanding that it is a personal debt of the sister's.

This suggests an 'AI can't see gorillas' problem here in that, during an AI-human interaction, identification of relevant big-picture context that a human advisor could have helped with is also missed.

For only 100 GB, that's quite expensive storage. Compare for example Backblaze B2 at $7.20 / year / 100 GB. That is just the storage, so if you do lots of I/O it might increase - but if you aren't using exactly 100 GB and don't do much IO it might also be less than $7.20.

I've personally been unable to pass AI 'liveness' detection (which was a high-stress situation when it related to something my new employer asked me to do after I already resigned from my previous role) despite repeated attempts and all I have is alopecia areata affecting my eyelashes / eyebrows (a relatively common condition).

These models are discriminatory for a lot of people, I'd say, and shouldn't be allowed.

Not even close. Israel apparently has AI bombing target intel & selection systems called Gospel and Lavender - https://www.theguardian.com/world/2024/apr/03/israel-gaza-ai.... Claims are these systems have a selectivity of 90% per bombing, and they were willing to bomb up to 20 civilians per person classified by the system as a Hamas member. So assuming that is true, 90% of the time, they kill one Hamas member, and up to 20 innocents. 10% of the time, they kill up to 21 innocents and no Hamas members.

Killing 20 innocents and one Hamas member is not a bug - it is callous, but that's a policy decision and the software working as intended. But when it is a false positive (10% of the time), due to inadequate / outdated data and inadequate models, that could reasonably classified as a bug - so all 21 deaths for each of those bombings would count as deaths caused by a bug. Apparently (at least earlier versions) of Gospel were trained on positive examples that mean someone is a member of Hamas, but not on negative examples; other problems could be due to, for example, insufficient data, and interpolation outside the valid range (e.g. using pre-war data about, e.g. how quickly cell phones are traded, or people movements, when behaviour is different post-war).

I'd therefore estimate that deaths due to classification errors from those systems is likely in the thousands (out of the 60k+ Palestinian deaths in the conflict). Therac-25's bugs caused 6 deaths for comparison.

It does work as long as the attesting authority doesn't allow issuing a new identity (before it expires) if the old one is lost.

You (Y) generate a keypair and send your public key to the the attesting authority A, and keep your private key. You get a certificate.

You visit site b.com, and it asks for your identity, so you hash b.com|yourprivatekey. You submit the hash to b.com, along with a ZKP that you possess a private key that makes the hash work out, and that the private key corresponds to the public key in the certificate, and that the certificate has a valid signature from A.

If you break the rules of b.com, b.com bans your hash. Also, they set a hard rate limit on how many requests per hash are allowed. You could technically sell your hash and proof, but a scraper would need to buy up lots of them to do scraping.

Now the downside is that if you go to A and say your private key was compromised, or you lost control of it - the answer has to be tough luck. In reality, the certificates would expire after a while, so you could get a new hash every 6 months or something (and circumvent the bans), and if you lost the key, you'd need to wait out the expiry. The alternative is a scheme where you and A share a secret key - but then they can calculate your hash and conspire with b.com to unmask you.

And once a CPU is attacked with a voltage glitching type attack, the compromise is so complete that the secret seeds burned into the hardware are leaked.

Once they are leaked, there is no going back for that secret seed - i.e. that physical CPU. And this attack is entirely offline, so Intel doesn't know which CPUs have had their seeds leaked.

In other words, every time there is a vulnerability like this, no CPU affected can ever be trusted again for attestation purposes. That is rather impractical - so I'd consider even if you trust Intel (unlikely if you consider a government that can coerce Intel to be part of your threat model), SGX provides rather a weak guarantee against well-resourced adversaries (such as the US government).

but not actually write it down or link the two identities digitally

What is to stop you just selling the ID card with zero consequences? Unless it has a photo on it of course, in which case that itself is an identifier you can't easily rotate.

Better is to use zero-knowledge cryptography to prove that you have a real ID's private key in your possession. Leaking the private key would be the same as giving away your real identity. Now you could make a proxy service that generates the proofs for money without it being traced back to you - but maybe a countermeasure to limit but not eradicate abuse would be for the protocol to include a proof you haven't used the same real identity to prove your age on that service in at least x days (that does mean you could be tracked for x days until you prove your age under another pseudonym).

At play here is the difference between necessary and sufficient conditions for democracy to be genuine.

There are a vast number of ways in which people can be denied a legitimate opportunity to obtain the policies most want (making a democracy less legitimate), but to list a few:

* People are allowed to vote, but votes aren't counted fairly.

* The voting system can be vulnerable to vote splitting (e.g. First Past the Post) or have other vulnerabilities that can lead to an outcome where the majority would have supported candidate A over candidate B, but B wins. This can lead to dynamics where tactical voting means only the candidates endorsed by the two best known parties have a chance, even if both of those parties are captured by minority interests. Generally the duopoly perpetuate the poor voting system to protect their interests.

* The leaders of any movements challenging the incumbents can be targeted before they have a chance to run - e.g. with trumped up legal charges, defamatory claims, violence either blatantly from the government, or by supporters which goes unpunished / is pardoned.

* Voters fear that if they vote against the incumbent, they'll face consequences.

* Voter demographics considered less likely to support the incumbent face barriers to vote, such as demands to provide documentation they might not have.

* Special case: Voting is restricted to citizens, but there are populations of people who live long term in the country but don't get a vote. Citizenship is not granted to or even stripped from people who are considered less likely to support the incumbent.

* Media is state owned, and is biased towards the incumbent, preventing the public from learning about alternative policy platforms in a meaningful way.

* Media is privately owned, and biased towards the interests of the owners of the media, preventing the public from learning about policy platforms opposed to the owner in a meaningful way.

* There are significant barriers to becoming a candidate (financial, or requiring a lot of work which costs a lot of money), preventing non-wealthy groups from being able to run.

* Corporations or ultra-wealthy are allowed to selectively fund large amounts of money (beyond the means of normal citizens), allowing policy platforms they support to drown out policy platforms in the interest of the public.

* As you mentioned, electorate boundaries are set in an unfair way (gerrymandering).

There are a lot more - but the summary is that there are many ways to undermine a democracy, and there are many countries that are nominally democratic, but aren't really.

If allowed, an incumbent monopoly will do the minimum innovation or tactical pricing necessary to keep out competitors - even if it means they suffer a temporary loss.

This can effectively create a barrier to entry high enough that no small company has a chance to beat them. Since insolvency before overcoming the barrier is foreseeable, no one even tries, and the monopoly gets to keep high prices (compared to a competitive market if they weren't there).

It depends what the proof system is being used for. There are applications of ZKC that this would absolutely break.

It isn't on the exact same proving scheme broken in this research, but consider https://risczero.com/blog/zkpoex, which is about proving that you have an exploit (a program) that puts a protocol into into an unexpected state, without revealing the exploit. Imagine you had a specially crafted program that allows you to prove you have an exploit, but actually none exists, and it is just that you're computing the same hash in your program that is used in the Fiat-Shamir heuristic, and violating the assumptions of the random oracle model.