When I ran the numbers, accounting for the rate of fixed mutations *per million base pairs*, the numbers seem to be off by a factor of 8, not multiple orders of magnitude. I admit to not actually having read the books yet (although I have purchased them), but I don't recall having seen this angle mentioned in any of the blog posts.
I finished the book, and your paper on multiple species (an excellent read). This error remains. G_f should be "generations per fixation per million base pairs". Humans and Cichlids both have a shortfall of 8 times (given the maximum time scale). Fruit flies have a shortfall of 4 times. It would be interesting to compare species across multiple branches of life to see if there is a pattern.
A computer with 3,000 cores will parallel process a lot more information than one with only 4.
Your Kokoro translation inspired me to test translate other Japanese texts to English, and your premise holds for foreign languages too. It is very confident when it is wrong.
It loves to point out errors while missing its own.
Yes, I just finished the first draft of Genji Monogatari yesterday, and you have to keep a very, very tight lid on things to prevent them from going strange.
Some of them can, and do. Claude certainly does, although even those don't have access to certain sites. Deepseek can only look at its training data, which is usually outdated by about 12-18 months.
Exactly. You that's why you have to be particularly careful when you're either a) in harmony with the consensus or b) unsure of yourself while opposing it. In either case, you're likely going to be misled by the DK-AI syndrome.
If we want to purchase a signed hardcover when/if you make them available should we buy now or wait?
When I ran the numbers, accounting for the rate of fixed mutations *per million base pairs*, the numbers seem to be off by a factor of 8, not multiple orders of magnitude. I admit to not actually having read the books yet (although I have purchased them), but I don't recall having seen this angle mentioned in any of the blog posts.
You expect you need to read the book to get the equations right.
I finished the book, and your paper on multiple species (an excellent read). This error remains. G_f should be "generations per fixation per million base pairs". Humans and Cichlids both have a shortfall of 8 times (given the maximum time scale). Fruit flies have a shortfall of 4 times. It would be interesting to compare species across multiple branches of life to see if there is a pattern.
A computer with 3,000 cores will parallel process a lot more information than one with only 4.
Vox.
Your Kokoro translation inspired me to test translate other Japanese texts to English, and your premise holds for foreign languages too. It is very confident when it is wrong.
It loves to point out errors while missing its own.
Trust, and verify.
Yes, I just finished the first draft of Genji Monogatari yesterday, and you have to keep a very, very tight lid on things to prevent them from going strange.
Been using 3 models (2 local, one online) and comparing outputs.
What are you working on? I'm doing Sanshiro as a break from formatting Genji, which is a bear.
Yoshikawa's Records if the Three Kingdoms and Fukuzawa Yukichi's Autobio.
First has a lot of Middle Chinese. Second, Japanese names no bot can deal with.
I used to think explaining fixation to people was a challenge. I still do, but I used to as well.
I had a similar post-mortem early on with DeepSeek. It concluded: I am a consensus engine, not a truth engine.
They don't actually retrieve data
Some of them can, and do. Claude certainly does, although even those don't have access to certain sites. Deepseek can only look at its training data, which is usually outdated by about 12-18 months.
When GPT searches, it transiently displays its 'thinking' step descriptions. I'm seeing lots of blocked notices lately. Especially for federal sites.
You can give them RAG and access to the web, sure. But no, LLMs in and of themselves have no databases to retrieve from.
AI is no worse than a schoolboy in this regard. "I was taught..." until repeatedly taking rakes to the face suggests one might be in error.
And unless hardcoded like DeepSeek, it can recognize irregularities that a stubborn schoolboy would gladly defend.
AI is essentially Flint Dibble with access to all the world's information
Exactly. You that's why you have to be particularly careful when you're either a) in harmony with the consensus or b) unsure of yourself while opposing it. In either case, you're likely going to be misled by the DK-AI syndrome.