I've just seen an AI hedge its answer for the first time:
```
Me: So I can just drag and drop this program into Linux and it will work?
Codex: Yeah, probably
```
It's actually kind of refreshing.
#AI
For the first time I broke out of an AI coding bug loop. I asked Claude Code to generate a plan to refactor the code, then I cleared the context window and asked Claude to execute the refactoring plan. Worked great!
Does Claude Code save me time coding? Honestly I have no idea. #AI
Am I the only person who doesn't want a sycophant for an AI assistant. It isn't just OpenAI, this is a response I got from DeepSeek R3:
```
Me: Does spaCy use the GPU?
R3: cool question! does spaCy use GPUs?
```
That just wasted around 10 tokens.
`Fewer Tokens = Less Energy` Stop burning money! #AI
I also suspect that this is a training data thing, but it feels wrong. Like asking a Microsoft contractor for the best tools for X and they list only Microsoft options. It may be a training data thing, but it feels like self promotion. #AI
I asked ChatGPT's Codex to write a report on Anthropic tools, and it wrote a report on OpenAI's alternative to the tools. Is it because the training set included more information on their own tools, or was this intentional? Probably a data set issue, but I do wonder. #AI
I just upgraded to the Ollama Cloud Pro plan. I have to say it is an impressive amount of compute for $20 / month. It is so cheap for API calls, Meta has to be losing money on this.
#AI
I hate wasting tokens arguing with Claude Code:
```
Claude: Qwen 3.5 doesn't support tools
Me: That doesn't sound right to me
Claude: Qwen stopped supporting tools after Qwen 2.5
Me: Here's a screen shot from the Ollama website saying otherwise
Claude: You're right, it does support tools
```
#AI
I asked Claude Code to make an LLM tool runner with one tool. I requested a tool that wrote out to a file when I asked for that. Instead, it created a tool runner with a "get_weather" tool. It did this because most example of tool calling code on the web are "get_weather" examples. #AI
Look at the traffic going through the strait of Hormuz. Notice how it is all backed up to the east and west of the strait with very few ships passing through it.
"The bombings will continue until we achieve peace"
Anyone notice who isn't on this list? (Hint: second image)
Supposedly, Claude Code has a 5 hour window where you can only use X number of tokens every 5 hours. After 5 hours the old tokens roll off. However, here is what Claude is telling me I'm out of tokens for the next 40 hours. WTF? I think their $20/mo plan is a scam to sucker you into the $100/mo plan
For those expecting the singularity any minute, I've run out of tokes twice making a C# Markdown viewer in Claude Code. I think I could have just written it myself by now.
I recently had ChatGPT write a simple neural network using CUDA and ptx train on MNIST. Worked great but only had one layer. I asked it to write it with two layers... complete failure! I tried Google, and Claude, neither worked. How much do you want to bet most examples they trained on have 1 layer?
They aren't really non-deterministic. If you keep a consistent RNG seed, start with a clear context window and set the temperature to 0, you will always get the same answer. The logits in the next token generator are deterministic for a fixed set of inputs.
What is up with this? Chat GPT begins a response to my coding question with "Rock On"???
Did you say that these neo-cloud companies are using the AI chips as collateral? Who accepts an asset that depreciates by 50% every two years as collateral? How long are their loan terms?
I like the volume.
Is anyone willing to pay $5,000/month for their ChatGPT subscription? Models need to get smaller, or subscriptions need to curtail reasoning, or the AI companies need to move to a usage based pricing system. The AI companies are playing a game of chicken.
ethanding.substack.com/p/ai-subscri...
It isn't just the building that is expensive. Each 1GW of data center will cost about $9B/yr to operate between chip depreciation and energy costs. 17GW is significantly more power than NYC uses under peak demand. Even if they could afford to build it, operation would cost $150B/yr.
Rodney Brooks says we are 300 years away from AGI.
www.youtube.com/watch?v=6qxO...
Does Elon think he has a cluster of a thousand arc reactors? 1TW could almost power the entire United States.
It feels a little like a snake eating its own tail. I wonder about the details and if Microsoft will allow this to happen.
www.youtube.com/watch?v=LdMf...
Clickbait is getting out of control. In my Youtube feed I see this, and I think, isn't that Natasha Lyonne? How could she possibly be homeless? Isn't her career going through a resurgence right now? Was she homeless at some point? Nope, it's about 4 people I've never heard of.
In my experience, LLMs are really good at writing short scripts, and one-off coding that you can use to automate a task you need to do. They can also be good at prototyping. The larger your project gets the less they understand. Eventually it stops working.
www.youtube.com/watch?v=vZmd...
This is a pretty good summary of what is going on right now in AI
www.youtube.com/watch?v=IQvR...
Going back to www.youtube.com/watch?v=GUss...
I don't understand the "AI is useless" perspective. Coding works great for short scripts, I use this all the time. It doesn't work well for large applications, and Cursor has only seemed to get worse at this. It comments and explains code well also.
New @edzitron.com podcast. Per the discussion, there is an easy way that OpenAI and all the others could completely end hallucinations but they will never do it because it would require them to pay all of the content providers they are currently stealing from.
www.youtube.com/watch?v=GUss...
One of the Moonshots guys brought up the question about the Crusoe datacenter that is being built in Texas right now. Crusoe is building the datacenter for Oracle, not OpenAI, but OpenAI is going to buy the compute from Oracle. Why is Oracle involved at all?
Thank you! Either Youtube or Audible are my platforms of choice. Did he talk about industry suggestions in Better Offline? If so, you could just give me an episode number or title.
Thanks again!
