Hacker News: matusp

New comment by matusp in "Continuous Diffusion Language Models (CDLM's)"

matusp — Mon, 31 Aug 2026 05:26:53 +0000

> I disagree with this. Decoders were absolutely dominant in 2020 for chat. GPT2 was considered too dangerous to release, and I remember scrambling to get on the GPT3 waitlist. It worked.

He's not talking about decoders, he's talking about auto-regression. Before ChatGPT, the dominant paradigm was fine-tuning BERT-like models.

> Before ChatGPT there really wasn’t much of a concept of pre-training and post-training.

Again, people spend years just post-training BERTs in various ways.

New comment by matusp in "DiffusionGemma Technical Report"

matusp — Fri, 21 Aug 2026 05:52:13 +0000

There is no theory why the gap shouldn't be closed. At the same time, people are trying to make them work for years by now and it's never good enough. But they are competing with incredibly optimized architectures.

New comment by matusp in "Patterns and problems in emerging multi-agent systems"

matusp — Sun, 16 Aug 2026 14:08:26 +0000

To me, all these agent systems just look very stochastic. You have these agents that have some basic computer capabilities and they are producing semi-random actions that also affect the semi-random actions of other agents. It is funny to observe how this stochastic system works, but it does not seem very practical to me so far.

The recent OAI-HF hack seems very similar. You have bunch of random actors and eventually they by chance iterated to a series of actions that breached HF environment. I don't perceive this as a malignant artificial intelligence, I perceive this as dangerous stochastic system that can control buttons that can affect the outside world.

New comment by matusp in "The Iran War Is a Whole New Level of Quagmire for the US"

matusp — Sun, 16 Aug 2026 07:29:09 +0000

The transformation from Roman republic to Roman empire was mainly about how the top-level government is organized. It was not about the relations with vassals/periphery. The Roman republic was as imperialistic as empire in this regard. In fact, most of the Roman conquering was already done when the republic ended.

New comment by matusp in "How Claude's text watermarking works"

matusp — Sat, 15 Aug 2026 08:54:55 +0000

I don't know the details of their implementation, but in general, if the text is not watermarked, you will probably fail the test after a few tokens. If it is, you will have to run the entire text, or, you can just run it until you have some degree of confidence.

New comment by matusp in "No one can afford to make Myst games anymore"

matusp — Sat, 08 Aug 2026 14:50:33 +0000

And when you are an investor, you must predict who is able to make a good game. What is the likelihood that a Myst-like will pay for itself. There is already a lot of Myst-likes on Steam, do players want more?

New comment by matusp in "“Gravity is worth asking about”"

matusp — Wed, 05 Aug 2026 07:07:48 +0000

The minimalist approach is great until it isn't. You can cover 98% of use cases with a single button is a great motto, until you are in the other 2%. I often fail to find basic functionality in software today, because some smart product manager decided that nobody really needs it. For example, it is now often impossible to load all elements from a list and filter/sort it along basic properties.

New comment by matusp in "What Emily Bender meant by "stochastic parrots""

matusp — Mon, 06 Jul 2026 17:24:22 +0000

It is a good blog, not a good paper.

New comment by matusp in "Exapunks (2018)"

matusp — Thu, 02 Jul 2026 23:21:05 +0000

The thing is, you can play it in any way you like. You don't have to optimize for speed at all if you don't enjoy it

New comment by matusp in "AI has already killed academia as we know it?"

matusp — Tue, 23 Jun 2026 05:37:07 +0000

With the volume of outputs in today's academia, this is simply not possible. There are conferences with tens of thousands of submitted papers, grants have hundreds of pages, etc.

New comment by matusp in "OpenAI Losses Increased Nearly 8X in 2025, with Spending Hitting $34B"

matusp — Tue, 16 Jun 2026 06:33:54 +0000

The AI companies also have a lot of space to grow their income (more ads, price hikes, ...). It seems realistic for them to turn profitable. But the market expected much more from these companies.

New comment by matusp in "Rome Fell and Nobody Noticed"

matusp — Mon, 15 Jun 2026 04:36:30 +0000

The standard of living in Rome was not surpassed for more than a millennium. Large parts of the empire were at peace for centuries, allowing economy and trade to thrive. Many provinces that were peaceful and prosperous were ravaged by wars, ethnic conflicts, and plagues for many centuries after the fall.

New comment by matusp in "How LLMs work"

matusp — Sat, 06 Jun 2026 07:11:59 +0000

There are thousands of people working in top level labs. Somebody would leak it

New comment by matusp in "They’re made out of weights"

matusp — Thu, 04 Jun 2026 08:34:47 +0000

Is matter code? There is some sort of computation happening in space over time.

New comment by matusp in "Can You Stop a Hypersonic Missile?"

matusp — Tue, 02 Jun 2026 04:02:27 +0000

I understand that motive. On the other hand, LLM smell makes the text untrustworthy. I have detected it as well, and I immediately started to wonder about whether I am reading a reasonable expert analysis or just an AI hallucination. I still don't know.

I recommend prompting the LLM to mostly fix glaring grammatical and stylistic mistakes, not to rewrite the entire thing into a LinkedIn post style text.

New comment by matusp in "The dangerous delusion of modern warfare"

matusp — Sun, 31 May 2026 14:02:15 +0000

There are many ethnic Russians in Eastern Ukraine that really welcomed the Russian army. The problem is that the rest of the country is not that happy about this.

New comment by matusp in "Anthropic surpasses OpenAI to become most valuable AI startup"

matusp — Sat, 30 May 2026 16:43:02 +0000

Can you give me some examples of these interactions / vibe?

New comment by matusp in "Leo's first encyclical attacks technological messianism"

matusp — Sat, 30 May 2026 13:31:33 +0000

This is the question that led to the communist movement in the 19th century.

New comment by matusp in "The Kaiser and a "Mediocre Man" Theory of History"

matusp — Sat, 30 May 2026 10:42:11 +0000

Of course mediocre and bad leaders make their mark on history. But Carlyle's Great Men theory is more about paradigm shifts that Great Men can will into existence, not just random noise they bring along. The problem with GM theory is that there is only a handful of examples to support it. Napoleon is one such example, and it was undoubtedly the inspiration why the theory was proposed in the first place. People were trying to come to terms with the fact that one leader can have such a dramatic impact on the entire world.

New comment by matusp in "The spread of Christianity, from antiquity until today, on an animated map"

matusp — Fri, 22 May 2026 18:48:37 +0000

I find it interesting how slow the spread really was after the initial burst. I am used to think about Christianity as a global religion. But before 16th century it was a pretty regional thing and its position seemed pretty precarious in handful of moments.