Carlos Baquero

@antirez Did you read the new @LynAldenContact novel? I also found it very interesting, on a topic not dissimilar, but don’t want to do spoilers.

In my novel Wohpe one chapter is the pope encyclical about AI. The novel is from 2022 (pre chatgpt). Here we are in 2026 with the actual pope writing an encyclical about AI.

I agree, it is cheating in a away :) and combining outputs of different models will explore diversity. Even with identical models I have been finding that fusing latent spaces is hard, and for different models it might even be impossible and require a natural language layer. Now the other research question is: What is the best common language for models to communicate? It might not be english, and I suspect a language can be evolved for more efficient model to model communication.

@xmal Yes but the cheat... :D all agents are the same model. The interesting part of putting things together is IMHO to exploit diversity. And a translation layer in this case must be trained in a complex way: at that point I bet that natural language becomes competitive.

Distributing LLM inference in DwarfStar: antirez.com/news/167

@antirez Not surprisingly they find in the latentMAS paper that latent space has more information than the information later produced in tokens, so it improves to comunicate that and avoid linking by a token creation step. Still, it is much more data intensive.

@xmal I was converging every hour more on the multi-agent paradigm indeed. Probably just using the natural language to make two models collaborate is a stronger candidate compared to continuation / token level ensemble. Reading. Thx!

Are you the same person at 55 that you were at 20? No. Your personality changes over time. •People tend to become more honest and humble with time. I remember struggling to admit that I did not know something that I should have known when I was in my 20s. Today, I am much more likely to just say, “Yeah, I did not know this.” Part of it, I think, has to do with the fact that I have built confidence. I have had time to show that I know things… I will readily admit, for example, that I am not an expert at any given technology in software. Though I have written programming books, I learn new idioms every week. I describe myself as a polyglot. I know a lot about a different things, but you can easily catch me having missed something. •Extraversion and agreeableness tend to increase with age. Though I still speak my mind, I am somewhat less direct today. When I first became a professor, I would grade assignments with comments such as “this is nonsense.” I have learned to be more indirect. I have relearned to be more direct at times in the last few years. I was a bit too mellow. •You become less emotional with age. I remember being mightily anxious as a young man. I would feel the anxiety. I rarely do these days. If I am trying to meet a deadline, I will simply work as hard as I can without thinking about the pressure. I am still stressed at times, I just don't think about it that much. Maybe it is just that I got used to it. •Openness tends to remain stable. I am as excited today about new technologies as I was when I was a young man. Because I am acutely aware of the dangers of “technological aging,” I always upgrade my skills. I learn new things every week. This makes it sound like aging is a great thing... and I do think that as far as personality goes, people do get somewhat better with time.

The digital divide has reversed. Digital is cheap, ubiquitous, often fake. Physical is the premium product now.

Fourth Europe Summer School on 
Distributed and Replicated Environments (DARE 2026) August 31 to September 4 | Milan | Italy

🇲🇻🇮🇹 FLASH | Les plongeurs finlandais qui ont remonté les corps pensent avoir identifié ce qui aurait provoqué la mort des cinq Italiens piégés dans une grotte sous-marine aux Maldives. Selon les premiers éléments, le groupe expérimenté aurait pourtant emprunté par erreur un mauvais tunnel lors de la remontée après avoir été trompé par une « illusion de mur de sable » dans l’obscurité de la grotte. Les secours ont retrouvé les corps dans un couloir sans issue situé à environ 50 mètres de profondeur. Les experts estiment qu’avec leurs bouteilles standards de 12 litres, les plongeurs ne disposaient que de quelques minutes d’autonomie à cette profondeur. Pris de panique et désorientés, ils auraient rapidement consommé tout leur air. Les enquêteurs analysent désormais les GoPro et le matériel récupéré pour comprendre précisément ce qui s’est passé. (La Repubblica)

I concur with the long term driver, adoption, and that in my view comes from the robustness and decentralization of the distributed protocol. Still, data is showing that at short term the dominant causality is from price towards on chain metrics, with hash rate and difficulty reacting faster. Actually I always assumed it was in the reverse direction until seeing the data.

@xmal @NeilJacobs $BTC's long-term value is driven by adoption and network effects. $BTC is structurally driven by adoption, but cyclically led by price. Selling bc the chain looks quiet after price weakness mistakes the lagging indicator for the driver.

NEW ‼️ - BILLIONAIRE MARK CUBAN: I SOLD MOST OF MY BITCOIN. IT’S LOST THE PLOT.

Finding from our Bitcoin paper, this one was surprising as I assume otherwise: Do adoption metrics drive price, or does price drive adoption metrics? Using Granger causality and cross-correlations on first-differenced log series, we find the stronger direction is: price → on-chain metrics Price changes lead hash rate, addresses, transactions, difficulty, and UTXO count. The lead is modest, but consistent: on-chain activity tends to respond after price, with peak lags around 13–61 days. So in this data, Bitcoin adoption metrics look less like the engine of price and more like part of the feedback loop that follows price.

X is being inundated with no-effort posts, especially replies, that are AI-generated and add nothing but noise.

@reubenbond And X even has a mechanism to integrate automated posts, they get a special automated label, but it’s rarely used and most developers just pretend it’s a real user posting.

@david_eng_mba @NeilJacobs Would say that adoption is the driving force, but was surprised to find that causality is much stronger from price towards chain metrics.

@NeilJacobs This is why deep understanding of $BTC matters. Mark’s analysis is shallow because it misses the two forces that actually drive Bitcoin: adoption and network dynamics.

@adamwarski We need voluntary labels like @creativecommons to identify classes of AI generated code.

Maybe we should stop pretending that we carefully review AI-generated code. Not all software is equal. Some of it is critical, but the majority is not. What if we separated human-written & AI-generated code entirely? This might be at the system, project, or module level: some modules might be end-to-end AI-generated and maintained, with people involved in writing feature requests, reporting bugs, and providing usage feedback. They wouldn’t, however, ever touch or read the code. That would resolve the (human) code review problem. But some code is critical, or needs the extra innovation, say, a novel architecture, that humans excel at. That code could be deliberately developed by humans ONLY. No AIs allowed. Hard questions remain: how to identify "critical" software upfront? What if a piece of AI-generated software becomes critical?

However, this also happens with a synthetic power law with added noise. So it does not invalidate de power law statement.

Another visualization of the shift impact.

Some observations from our recent analysis of the BTC power law. The famous log-log price fit is real: Bitcoin’s price sits surprisingly close to a power law over 2010–2026. But the exponent is not robust. Change the time-origin shift and the fitted exponent moves from about 5.65 to 16.49. So the power law is visually strong, but structurally fragile.