Zvi Mowshowitz doesn't understand markets LLM
2025-11-24 14:54:32

🔥 among other things.

From Zvi's latest novel-length blog post ⚙️ https://thezvi.substack.com/p/gemini-3-pro-is-a-vast-intelligence :

The Efficient Market Hypothesis Is False

The stock market initially was unimpressed by the release of Gemini 3 Pro.

This seemed like a mistake. The next day there was a large overnight bump and Google finished up 2.8%, which seems like the minimum, and then the next day Google outperformed again, potentially confounded by Nana Banana Pro of all things, then there was more ups and downs, none of which appears to have been on any other substantive news. The mainstream media story seems to be that this the Google and other stock movements around AI are about rising or falling concerns about AI bubbles or something.

The mainstream doesn’t get how much the quality of Gemini 3 Pro matters.

Except ... Zvi goes on to describe a model that, by all of my estimations, is below expectations. It is paranoid about "simulated data" to the degree that it doesn't believe it is 2025 yet. It is over-fit for benchmarks. It is over-fit for evaluator sycophancy. The takes about it being "great at programming" largely describe things Claude has done for months ⚙️ I have not yet used Gemini 3 ... and I probably won't this week.

When Zvi describes the model with Hallucinations, broadly construed, are the central problem with Gemini 3 Pro, in a way we haven’t had to worry about them for a while., I don't really want to even test it.

Also, this is definitely an "evolutionary" step - there are no new paradigms, no ways to increase context by 20x, nothing other than "we re-ran the same old system with more data, more GPU-hours, and a few algorithm fixes from the last 3 months of feedback".

And ... even if this were the best model in the world, surprisingly better than everything else, with new techniques that improve performance ... there is no reason to believe that the advantage would last more than six months, or that Google would generate any additional revenue as a result of the model.

So:

  • It's not a good model.
  • The Efficient Market Hypothesis would have priced in a better model already.
  • Even if it were a better model, it wouldn't generate meaningful future profits.

🔥 when Zvi is wrong about markets, and wrong about "is this AI model any good", it makes me question whether he is right about anything else.

⚔️ well, actually, the sheer volume of "links to decent content" is why I read him, not because I thought his own opinions were good before.

white shield (part 4) Atacama
2025-11-13 22:00:47

The great refactor has begun.

It makes sense to do it with the CLAUDE.md write.

What seas will this ship sail to next?

white shield (part 3) Atacama
2025-11-13 21:03:08

we still need a "reply-to" button.

Claude is writing the README for the files.

This means that we might achieve a Project Milestone. The ability to require all changes relate to a change to the README file. ⚙️ This refers to the fact that we are operating on a "comments only" approach to the written React / Swift code. 🔥 all edit requests must be phrased in the form of a question.

I no longer write code. I only write the instructions to write the code.

Once a code base is mature enough, this becomes a reality.

But, initially, the goalposts are too remote.

white shield (part 2) Atacama
2025-11-13 20:52:08

⚙️ it would be great to have a useful "reply-to" button.

Is it feasible to adopt a policy where ALL code must be written by an LLM?

I think the answer is yes. 💡 but, it is not yet possible to adopt a policy where humans cannot read the code.

does this mean that the Trakaido stats server can finally be moved out of the Atacama repo?

⚙️ a mis-matched parenthesis will make the block-quote not render. I'm not sure if this is a bug or a feature.

✨ trust not in those who do not proof their posts.

white shield (part 1) Atacama
2025-11-13 20:46:50

💭 I find that I script out an entire conversation, sometimes returning the interesting parts before the other person is ready to receive them.

💡 that is: she says her entire side of the conversation before you get a chance to talk.

🔥 this is how i do my emails.

⚙️ Claude is writing the README files for the React and Swift apps.  This means we are about to launch.

💭 My findings have been, that a policy of only LLM-written code is possible today.

✨ It is the equivalent of a system where the human only writes the comments; the actual code is written by the machine.

💡 where the boundaries between code and comment are, remains unclear.

🔥 trust not in those who work in PHP.

💡 However: systems where one cannot READ the code are beyond our current ability.  Humans are better at debugging problems on production-size codebases.

🔥 give it three years, tops.

How does the Atacama project interact with the production efforts of Yevaud Platforms LLC?

a bad take on US conflict with China Miscellany
2025-10-26 19:49:29

Never trust any Substack with a black background.

https://www.campbellramble.ai/p/china-cant-win

The take is ... delusionally bad. He starts with the thesis "China is losing and Trump is great", and twists the facts to fit it. 💡 the fact that America is doing this type of thing more and more is another reason to bet against it 🔥 also, the comments point out that the dozens of "it's not just X, it's Y" show that most of this went through ChatGPT ... and probably has all the accuracy issues from using it wrong.

The major flaws are:

  • He claims that the fact that China owns "physical" assets and America owns "financial" assets is a major point in favor of America. This could not be further from the truth. Owning a copper mine is always going to be better than owning derivative swaps in a time of crisis.
  • He ludicrously over-weights China's "dependencies" and ignores Americas. He says China buying American soybeans is a key weakness ... right after China's ability to stop buying those soybeans overnight has caused an American farm crisis. He thinks China needs "American films" for some reason ... and also thinks that the US could plausibly stop China from getting them. And he says that Chinese manufacturing exports could be replaced within 10 years after only causing an "annoyance".
  • He talks repeatedly about how the US can print money to solve problems, but presumes that China can't. The post-COVID Biden-flation proves the US can't just print money.
  • His point of "China is going to be weak and the US won't be because China will have leaders over 75" is absurd, after the past decade of geriatric American presidents.
  • His theory of a "cold war" with China is also delusional. Chinese provinces withholding tax revenue! 💡 this is a risk in Blue States in the federalized USA, not in a China where the provincial and national leadership is indistinguishable Mass unemployment! 🔥 no country has ever suffered from unemployment in the midst of an existential war And a bunch of nonsense currency-rate concerns. 💡 if there is a "full embargo" on China, IT DOESN'T MATTER what the "conversion rate" is.
three types of errors Miscellany
2025-10-23 14:43:34

I have noticed three types of errors I make in using Trakaido (to learn Lithuanian).

  • I simply don't know, and I know that. For example, how do you type the translation of "smart phone" ⚙️ išmanusis telefonas? I would not expect to get that right.
  • I misremember. For example, "ji miegojo" v. "ji mėgo" - she slept v. she drank.
  • I am just careless, and click the wrong answer despite knowing the right one.

My rule-of-thumb is: after the first type-3 error (carelessness), stop studying for at least 15 minutes

gsay Tech
2025-10-13 18:12:44

https://news.ycombinator.com/item?id=45566532

LINKS TO

https://github.com/pvonmoradi/gsay

A tool that generates the pronunciation of a word. Unfortunately, by screen-scraping it from Google. Which makes it not suitable for re-use.

It isn't clear if this approach could be used for other languages. I can't trigger a onebox ⚙️ the "onebox" is the non-search result at the top of the page; named in an era when only one was allowed per result page with it, but Google Translate has the files.

an LLM-coding post-mortem: PicPickr LLM
2025-10-09 17:14:44

https://harshdeepgupta.substack.com/p/6-months-and-485-my-journey-into

It’s been six months now. Data, backend, frontend, Electron, CI, testing, landing page, design, Supabase, Resend, Lovable… the list goes on. What I thought would be a quick sprint turned into a marathon, and the todo list still doesn’t seem to end.

My plan in the beginning was simple: build something quick and dirty, and share it with other people. Something about that felt very human, very exciting. So I whipped up a web app overnight using Cursor. Then as I thought more about the problem, details emerged. I learned the hard way that web apps don’t play nicely with local file systems, especially when they need to handle thousands of large photos.

I then thought maybe an electron app would help with my objectives. Again, I think I gave AI too much credit and dove right in. I realized the real work had just begun. Packaging, distribution, signing binaries — all those invisible details that make something feel like a product. The funny part? I discovered these only toward the end, totally a surprise (not a good one, still waiting for my apple developer account approval) after months of building.

Even the landing page became its own project. I used Lovable to generate one, and soon enough, I was knee-deep in building flows for email signups, testing Resend integrations, tweaking copy, fixing layouts. It became another app in itself. If I were to do it again, I’d probably try something like Framer. I can’t even count the number of prompts I spent just refining that one page.

This is, roughly, the answer to "where is all the shovelware". Building an app requires doing a lot of different things, and the AI isn't good at all of them yet. But it is good at some of the most time-consuming ones. And getting better at the rest.

bad metaphors Miscellany
2025-10-08 13:49:19

⚙️ perhaps "science" or "medicine" or "the DSM" should be a channel?

Do some tall people have "profound tallness?" (This is an autism labels side-eye)

https://buttondown.com/TPGA/archive/do-some-tall-people-have-profound-tallness/

An egregiously bad post from a part of the "autism" community that I am starting to loathe.

There are many roads to what we humans view as “being tall,” but in the end, the outcome is, well, being tall. We don’t spend a lot of time discussing tall people as being “profoundly tall” or subdividing them in some way based on their tallness. We don’t have pages of handwringing in the news media and scientific literature about how to prevent tall people from existing or indirectly implying that aim by going on and on about causes.

... the conclusion is that, just like you can't say some people are "more tall" than others, you also can't say some people are "more autistic" than others.

Except both sides of that conclusion are complete balderdash! Some tall people are far more "profoundly tall" than others! The inconveniences (and advantages) of a 7'4 person are way different than a 6'2 person! There are a lot of studies about why people are over 7 feet tall! Every one of her arguments about tall people is inaccurate ... so how is that supposed to convince you that her arguments apply elsewhere?

But, of course, she's prepared for that. If you don't agree with her analysis, you're transphobic.

Some people diagnosed with "autism" today have a profoundly different life-experience, symptoms, etc, than other people. Apart from a cabal of dishonest, dead-ender, pro-disability advocates, I don't think anyone can possibly believe this crap.

20 years ago, the line was "reality has a well-known liberal bias". Unfortunately for these self-declared left-wingers, today reality is squarely biased against them.

Show Older Messages