Hacker News Reader

kadushkaon AI capex is so big that it's affecting economic statistics

I love it how people are transitioning from “LLMs can’t reason” to “LLMs can’t reliably reason”.

kadushkaon AI capex is so big that it's affecting economic statistics

Depends on how that wealth is distributed.

kadushkaon Anthropic tightens usage limits for Claude Code without telling users

I use o3 to brainstorm research problems, and it's been pretty useful. Especially the deep research feature.

A parent should provide their kids with opportunities to try new things. Sometimes this might require gently making a kid do something at least a few times until it's clear it's not something they are good at or interested in. Also deciding when to try something is important - kids might need to try it at different ages. And of course convincing and reassuring a kid might be necessary to try something they are afraid to do. Until the age of 12 or so, it's important to make it fun, at least initially.

It's debatable whether a parent always needs to "lead by example": for example, I've never played hockey, but I introduced my son to it, and he played for a while (until injuries made us reconsider and he stopped). For mental well-being, make sure to not display your worst emotions in front of your kids - they will definitely notice, and will probably carry it for the rest of their lives.

kadushkaon Reflections on OpenAI

Many people are bad parents. Many are bad at their jobs. Many at bad at both. At least this guy is good at his job, and can provide very well for his family.

kadushkaon Ask HN: How much of OpenAI code is written by AI?

At some point, if this process continues, it is expected to actually automate software developers, would you agree?

kadushkaon Anthropic signs a $200M deal with the Department of Defense

Living in US is voluntary.

kadushkaon Anthropic signs a $200M deal with the Department of Defense

This might makes sense if you believe US is an evil empire, DoD is doing bad things, and AI will help DoD do even worse things. But it's not so black and white, is it?

kadushkaon Anthropic signs a $200M deal with the Department of Defense

With self-driving cars some human will be held responsible in case of the accident, I hope. Why would it be different here? It seems like a responsibility problem, not a technology one.

kadushkaon Anthropic signs a $200M deal with the Department of Defense

If there's not enough money from taxes, they will borrow/print more to cover total deficit (not specific to DoD). Otherwise, tax money will go directly to DoD.

kadushkaon Anthropic signs a $200M deal with the Department of Defense

Why do you think humans would make better life or death decisions? Have we never had innocent civilians killed overseas by US military as a result of human error?

kadushkaon Anthropic signs a $200M deal with the Department of Defense

If you live in US, taxes you pay directly fund DoD. So if you sponsor their activities, why can't Anthropic do business with them? Which other company would you rather get their (your) money?

kadushkaon Ask HN: To what extend have you stopped or limited your use of AI?

I’m thinking about limiting the extent I use it for coding. Just to stay sharp. Need to exercise my brain more.

kadushkaon Grok 4 Launch [video]

Thank you! I guess if there's enough spelling related text in the dataset, a model is forced to learn some info about token composition in order to predict such texts.

I wonder if it would help to explicitly insert this info into an embedding vector, similar to how we encode word position info. For example, allocate the first 20 vector elements to represent ASCII codes of token's characters (in some normalized way).

kadushkaon Grok 4

Oh yes, you’re right, I was thinking about claude.ai

kadushkaon Grok 4

The $20 one doesn't have Opus

It does.

kadushkaon Grok 4 Launch [video]

As long as it solves my technical tasks, I don't care what political biases it has.

kadushkaon Grok 4 Launch [video]

Is a human who never learned how to read not generally intelligent?

kadushkaon Grok 4 Launch [video]

Is there any evidence to support your hypothesis?

kadushkaon I don't think AGI is right around the corner

Great, then we established that I (generally intelligent human), and LLMs, both can't perform the task in question without specific training. We still don't know if given that specific training we (a specific human or a specific LLM) would be able to perform it.

kadushkaon Grok 4 Launch [video]

I got 3 tokens: st, raw, and berry. My point still stands: processing "berry" as a single token does not allow the model to learn its spelling directly, the way human readers do. It still has to rely on an explicit mapping of the word "berry" to b e r r y explained in some text in the training dataset. If that explanation is not present in the training data, it cannot learn the spelling - in principle.

kadushkaon Mercury: Ultra-fast language models based on diffusion

o3 is definitely usable, as I said, it solved about half of the coding tasks I tried. My problem with your original comment was "bottlenecks in genAI is not the knowledge or accuracy". Knowledge and accuracy are absolutely the main bottlenecks for LLMs today. Hallucination rate for o3 and o4-mini models have doubled (compared to o1), and OpenAI does not understand why. If my AI model is not accurate, and if it makes up fake knowledge I don't care how fast it is - I will have to spend more time double checking its output than the time I saved by getting that output faster.

kadushkaon Grok 4 Launch [video]

Each character made the cut, but the word "strawberry" is a single token, and that single token is what the model gets as input. When humans read some text, they can see each individual character in the word "strawberry" everytime they see that word. LLMs don't see individual characters when they process input text containing the word "strawberry". They can only learn the spelling if some text explicitly maps "strawberry" to the sequence of characters s t r a w b e r r y. My guess is there are not enough of such mappings present in the training dataset for the model to learn it well.

kadushkaon Grok 4 Launch [video]

These models are not trained on character level input. Why would anyone expect them to perform well on character level puzzles?

kadushkaon Mercury: Ultra-fast language models based on diffusion

Codex paper confirms that GPT-3 could not do any coding tasks. It's right there in the abstract: https://arxiv.org/abs/2107.03374

kadushkaon Mercury: Ultra-fast language models based on diffusion

Because gpt-3 was not trained to do coding tasks. It could do a simple autocomplete. Perhaps you are confusing it with gpt-3.5?

kadushkaon Mercury: Ultra-fast language models based on diffusion

gpt-3 could not do any coding tasks.

kadushkaon I don't think AGI is right around the corner

Last time I checked they were trained to predict the next word. Why would you expect them to solve research problems?

kadushkaon Mercury: Ultra-fast language models based on diffusion

What did he do with gpt-3?

kadushkaon First-Time Home Buyers Are MIA. Landlords Are the Winners

Move somewhere cheap and work remotely.