This statement only makes sense when viewed in economic terms. If every player keeps training ever-larger models, any company that falls behind will end up with an inferior product and lose market share. As a result, everyone is forced to continue scaling up. However, AI model training has become so expensive that it can no longer be funded through revenue from AI services alone. This competitive race therefore leads all participants toward eventual bankruptcy.

The only economically sustainable solution is to slow the pace—which is exactly what Anthropic is proposing.

That said, this is a textbook example of the prisoner’s dilemma: If everyone pauses, everyone receives a modest reward. But if others pause while one player continues training, that player can become the dominant force, drive competitors out of the market, and reap a far greater reward. I bet the collective pause won’t happen—and that many major players will ultimately go bankrupt, potentially triggering a major economic crisis.

When E2EE was first rolled out, a message appeared in each chat saying that communications were now secure. I always wondered how they managed to distribute the keys without Facebook ever gaining access to them. I long suspected that they might secretly keep a copy of the keys, perhaps obtained during the key distribution process itself. Now those suspicions are gone

It's not the first graphic adventure, but it IS the best game in the world.

Because I enjoyed it so much. Because it brings me fond memories of my childhood. Because I would love to be a kid again to play it for the first time.

Others might be considering technical reasons, but these are the real ones.

This exploit seems to be very useful for task assignments. Given that many students rely on AIs to do their assignments for them, the teacher can send cover instructions in the first task assignment (which are recorded in the long-term memory), and in a subsequent assignment get a result that discloses the use of AI.

Actually, packet marking to indicate congestion is not the novelty here, this has been employed (without much success) in IP for a long time. The novelty is how to respond to those marks and how often to set them.

Traditional TCP congestion control deals with an ECN marked packet similarly to a dropped packet, and it drastically reduces its congestion window, drastically reducing transmission speed. For this reason, ECN bits are traditionally marked only when the network is really congested. In this approach, ECN is not much different than a dropped packet, and its improvement is marginal: it avoids some retransmisions, but performance drops significantly when the network is close to congestion.

The approach in L4S is to mark packets more frequently, and let the transport protocol (TCP, QUIC, etc) adapt smoothly to the proportion of received packets with ECN bits set. It is clearly explained in RFC 9331. This idea borrows from DataCenter TCP, which pioneered the approach in 2011 or so.

t was clear that the AI could think [...]

Sorry, how was this clear, exactly?

Thanks for your analysis! I am more skeptical, but I see all your points.

Thanks, your comment is very clear; of course valuation depends on future expectations. In fact, this is the source of bubbles (over-inflated expectations).

I see it very hard to materialize, so I would not personally invest my money given such a huge valuation. What is your bet?

$150B is the approximate market valuation of Morgan Stanley, Total Energies, Comcast and Goldman-Sachs. This valuation puts the company in the Top-100 companies in the world. These companies have a yearly revenue of $96B, $237B, $121B and $46B, respectively.

OpenAI's yearly revenue is estimated to be around $3.4B.

What did you mention about a bubble?

You are completely correct, but I checked the results from 2024 and it was won by USA. Best result since 2019, in which USA shared the first place with China.

Do you know what would be an incredible help? Designing cars that can comfortably carry 3 infants, with their baby seats. In most cars (at least in my country) having more than 2 baby seats is impossible, and this becomes an obstacle to have another child, which you won't be able to move with. I mean ordinary cars, not huge vans with 3 rows of seats that most low- and middle-class people cannot afford.

If the government wants to implement increased birthrate policies, they should make this compulsory, by law.

And do not forget that these alternatives do not give the company the right to access creators’ content and use it freely.

Any industrial activity requires energy. This is a good use for such energy. On the opposite side we have bitcoin mining, which is (mostly) wasting energy.

The headline seems to imply that this is a bad thing. Just build the factories and the (nuclear!) power plants, dammit.

From the actual (retracted) paper:
RECEIVED 17 November 2023
ACCEPTED 28 December 2023

Anything from this open-access publisher is crap. They accept crap papers, and they do it in an incredible short period to make authors happy. No serious review can be made in such short period, and it is really strange to have a paper accepted in the first round of review -- unless, of course, the review process is crap. Also note they do this so fast to charge the authors with the publishing fees (yes, authors pay to have their work published instead of being paid; this is the current trend in scientific publishing).

They're breaking the game -- for nothing