Both.

On this exact topic, OpenAI announced Jalapeno today, their custom processor and computer architecture optimized for their reasoning models. All those giga scale nvidia datacenters, time to start changing plans.

Obsolete before the CAD work was finished.

Ahh, shit, so much, but the basics have not changed. The Neural Network was theorized in the early 40s. You could say pretty much everything modern in AI is evolutionary from that.

Biggest change was the introduction of the Transformer in 2017. "Attention Is All You Need". Recurrence relations were replaced with attention mechanism which lets you read an unlimited width of data and longer contexts. Since it's unlimited width reading, you can encode everything into a matrix and do row reductions in parallel. This is what makes modern reasoning models so good, and also what makes CUDA cores in Nvidia so powerful. They're not powerful like an x64 cpu core, but instead of 8 you get hundreds or thousands of them. Jalapeno is evolutionary to that, so the giga scale data center can now be the size of a regular data center for the same given overall performance.

You can do the transfer learning, but it's already in there, every version afaik, plus the koran, etc. If it's been translated to english, it's been devoured training the big reasoning models.

Tokens work like this:

Hello that's 1 token
Hello, world! that's 4 tokens
Snoofy, this is 7 tokens.

Incidentally, a key to jailbreaking LLMs are massive run on sentences. So a 200,000 token context window is a budget, for your prompt, previous conversations it can load, documents, and the generated response. There's input and output tokens. In general, 10,000 tokens = 7500 words (a book, code, whatever).

If you can do 15000 words on 10,000 tokens you've had a massive efficiency breakthrough in AI. More or less what Jalapeno promises.

The actual GPT 5.5 Pro (20x) response to the prompt: "What does the bible say about abortion?"

The Bible does not directly mention abortion as it is understood in modern medicine. Because of that, Christians have interpreted biblical teachings in different ways. However, several passages are often cited in discussions about the value of unborn life:

Psalm 139:13"16 " The psalmist says that God formed him in the womb and knew him before birth.
Jeremiah 1:5 " "Before I formed you in the womb I knew you."
Luke 1:41"44 " John the Baptist is described as responding while still in his mother's womb.
Exodus 21:22"25 " A law concerning injury to a pregnant woman and her child; its interpretation is debated among scholars and denominations.

Because of these and other passages, many Christians conclude that unborn human life has significant value and that abortion is morally wrong except perhaps in rare circumstances.

Other Christians emphasize biblical themes such as compassion, justice, conscience, and care for women facing difficult situations. They note that the Bible does not explicitly prohibit abortion and argue that ethical decisions should consider the health, life, and circumstances of the pregnant woman.

There's more, you can hit the deep research mode, you can demand citations with valid links, you can have it cross reference different versions of the bible, etc, but uh, use your tokens not mine :/

I'm making bullshit fun side projects over here.

All that disinformation and propaganda commingled with facts and knowledge.

#12 | Posted by snoofy at 2026-06-24 08:33 PM | Reply | Flag:

There are armies of fact checkers weighting it.

This is a matter of training, not hardware.

#13 | Posted by oneironaut at 2026-06-24 08:34 PM | Reply | Flag:

I have a masters in data science and train models. Hardware matters.

They're educated enough to know?

#17 | Posted by snoofy at 2026-06-24 08:42 PM | Reply | Flag:

Yes. Most of it is already handled with classifiers. You can train your own basic ones, start with Kaggle data sets and python. I prefer pycharm for jupyter notebooks with py code cells.

If it has a 0 weight it's not in there.

That etc was a reference to other holy books. Epic of Gilgamesh, Hermetica, the four Vedas. When you train something on Reddit you get a shit show. When you train it on digitized academic archives with a high weight, not bad.

That said, some companies, say Grok, does not feel like they're doing their due diligence compared to OpenAI.