lluisantoni

If you work with voice agents, I think this is an interesting paper. The figure I would like to highlight is Figure 1. Basically, the distribution of times between turns in conversations is unimodal and the mode is somewhere between 0 and 200ms, in any of the 10 languages they show.

I assume you are already using CLAUDE.md to give these instructions and still not working?

The placebo effect of useless dashbords.

That is a good point. And this issue looks analogous to what is happening in the research publication world (arxiv, peer-review, etc). In particular, I wonder whether code pushed is of lesser quality, CI pipelines break more often too, and this saturates github servers.

You should go for it, especially because nowadays Claude is a powerful tool available to many. One piece of advise though, everything Claude mentions that you do not understand or not know, make sure you learn it. This is probably the most important part, and applies to everyone using Claude.

In my opinion, the suggestions mentioned are great. In particular: 1) starting early or working late, and 2) pick 2 or 3 time slots a week to mark in your calendar just for you (to go out for a run, pick up kids, etc). The important part is that you do not skip these scheduled breaks for the sake of work! While it is hard, most of it is being mentally able to split work and personal life.

Not sure if someone mentioned it already but I also used bitbucket in the past and found it very easy to use with a similar offering as github.

Text watermarking is another EU rule made without real world input. The Union is stuck on major economic crises (electricity prices for instance) because nobody can agree on anything. However, the bureaucracy forces tech into a privacy nightmare. Brussels cannot bring together its own members but it loves pretending it can govern the internet.

Also, the authors claim the starch "inhibited lipid absorption." They do not argue why they used surrogate blood biomarkers instead of residual energy output in fecal samples (which is a better measurement).

sample size is 37, number of authors 29

For some companies there might be a need to run them locally. For instance, Apple decided to run LLMs on the phone locally. I guess it depends on how important latency and privacy are. Perhaps Meta is looking at how much interest for those local models is there.

I didn't think about this happening. It reminds me of the cross-border challenges with crypto. Each country has laws to control its author rights or money supply, but those are hard to enforce in the international setup.

This should not stop companies from doing it though, right? For instance, in autonomous cars reliability is key for human safety and we already have driverless taxis.

I found this post very good in an expected way. I still keep an old metro card from when I moved out of the city. Now can get the full vending machine.

The Sony step is also interesting. I have noticed that digital games keep the original price even after a few years while the physical ones, if you search for a bit, you can always find them discounted (the reason may be physical inventory management?). This is excluding memberships. The important point is that prices are probably better behaved if we include stores as a third party sellers, instead of having Sony the only player controlling prices.

ArXiv is not peer reviewed. When looking into more curated ML articles, I prefer looking into NeurIPS, ICML, or ICLR proceedings.

There is a Chinese parable about luck: "The old man lost his horse" https://en.wikipedia.org/wiki/The_old_man_lost_his_horse

Not sure how useful this is but offers a different perspective (and stoic answer) into your question.

Also, something else to add. At first I thought no one wants to build data centers in hotter areas in the middle of deserts (many places in the American continent). So nobody would spend money building a data center in the Chihuahuan Desert for instance. However, a game of latencies will either require cover llm access from these areas, or make people move closer to the other data centers. In the former, llm prices will go up; in the latter, there will be a migration towards data centers that increase the price of the areas around.

There is something I have been pondering recently. If we compare the cost of AI subscriptions (let's say Claude's 100/month) to a median developer salary (let's say 100k/year to 200k/year), the difference is orders of magnitude. This fills like a gap that needs to close. I suspect llms are too cheap right now but will raise their prices to a point where only big companies will be able to afford subscriptions to use them. I think soon we will see models that are only sold at very high prices.

I really appreciate the simple design of the tool. Could you share a bit about what happens with uploaded images, and how can such a tool survive without accounts (advertising)?

It feels like governments might not be as stable as a company to fund large scale research given how many years can a government last. I have discussed with a friend at Argonne National Laboratory how quickly projects can die and management change based on ruling party. That being said, there are some topics the US government keeps funding no matter the party.