290 karma · joined November 16, 2021
1. GPUs create heat. There's no efficient way to get rid of the heat in space (vacuum is an insulator). 2. Die-shrink makes modern processors and memory more and more susceptible to radiation; shielding is possible, but adds cost + mass (which adds cost)
Also (shameless self-promo) I publish a 2x weekly blog just to force myself to keep up: https://aimlbling-about.ninerealmlabs.com/treadmill/
If you can forgive the obviously-AI-generated writing, [CPUs Aren't Dead](https://seqpu.com/CPUsArentDead) makes an interesting point on AI progress: Google's latest, smallest Gemma model (Gemma 4 E2B), which can run on a cell phone, outperforms GPT-3.5-turbo. Granted, this factoid is based on `MT-Bench` performance, a benchmark from 2023 which I assume to be both fully saturated and leaked into the training data for modern LLMs. However, cross-referencing [Artificial Analysis' Intelligence Index](https://artificialanalysis.ai/models?models=gemma-4-e2b-non-...) suggests that indeed the latest 2B open-weights models are capable of matching or beating 175B models from 3-4 years ago. Perhaps more impressive, [Gemma 4 E4B matches or beats GPT-4o](https://artificialanalysis.ai/models?models=gemma-4-e4b%2Cge...) on many benchmarks.
If this trend continues, perhaps we'll have the capabilities of today's best models available to reasonably run on our laptops!
1. In my own use, since 1 Apr this month, very heavy coding:
> 472.8K Input Tokens +299.3M cached > 2.2M Output Tokens
My workloads generate ~5x more output than input, and output tokens cost 5x more per token... output dominates my bill at roughly 25x the cost of input. (Even more so when you consider cache hits!) If Opus 4.7 was more efficient with reasoning (and thus output), I'd likely save considerable money (were I paying per-token).
2. Anthropic's benchmarks DO show strictly-better (granted they are Anthropic's benchmarks, so salt may be needed) https://www.anthropic.com/_next/image?url=https%3A%2F%2Fwww-...
https://www.anthropic.com/_next/image?url=https%3A%2F%2Fwww-...
The final calculation assumes that Opus 4.7 uses the exact same trajectory + reasoning output as Opus 4.6. I have not verified, but I assume it not to be the case, given that Opus 4.7 on Low thinking is strictly better than Opus 4.6 on Medium, etc., etc.
Happy to take suggestions on this!
However, the essay and the guidelines were all human-written!
I honestly don't mind someone else's AI as long as I can trust it/them. One problem I have with sloppypasta specifically is that it reads as raw LLM output and the user isn't transparent about how they worked with the AI or what they verified. "ChatGPT says" isn't enough; for me to avoid inheriting a verification burden, I'd also need to understand what they were prompting for, if they iterated with the AI, and if/what/how they validated.
(the other problem is that dumping a multi-paragraph response in the midst of a chat thread is just obnoxious, but that's true even if its artisanal human-written text)
I've found success having sidebar conversations with the colleague (e.g., not in the main public thread where they pasted slop), explaining why it was disruptive and suggesting how they might alter their behavior. It may also be useful to see if you can propose or contribute to a broader policy on appropriate AI use/contribution with AI, and leverage that policy as the conversation justification?
sloppypasta: Verbatim LLM output copy-pasted at someone, unread, unrefined, and unrequested. From slop (low-quality AI-generated content) + copypasta (text copied and pasted, often as a meme, without critical thought). It is considered rude because it asks the recipient to do work the sender did not bother to do themselves.
'"epic vampiric doom" suggests a bright and uplifting soundscape'? I'm not so sure about that
... I was hoping for something more like nightwish.
I traced this idea across four layers of building with AI: how alignment training plants karmic seeds, why the intention behind a prompt shapes the model's trajectory, how memory turns an agent's first impressions into self-fulfilling karma, and what happens when no one tends the karma at all.
An excerpt:
> Alignment with human values requires not only understanding what "good" and "bad" actions are, but why they matter. Anthropic have planted karma through the intentional, considered design of the Constitution and its use in training Claude. It may manifest in a model that in some sense understands its own karma — alignment, it turns out, may just be karmic awareness.
I trained a bunch of simple linear regressions - while Omniscience Accuracy had the best fit (R2: 0.98), it predicted absurd multi‑trillion param sizes (Gemini 3 Pro ~1,254T total parameters). Artificial Analysis' Intelligence Index provided more plausible results:
Gemini 3 Pro: 3.4T Claude 4.5 Sonnet: 1.4T Claude 4.5 Opus: 4.1T GPT-5.x series in 2.9-5.3T range total parameters.
Interesting notes:
- task benchmarks (Tau²/GDPVal) aren't predictive of model size - adding price made the fit worse - sparsity or parameter activation ratios did not influence predicted sizes at all.
https://venturebeat.com/ai/why-do-87-of-data-science-project...
- I strongly recommend Chip Huyen's books ("Designing Machine Learning Systems" and "AI Engineering") and blog (https://huyenchip.com/blog/).
- Andreessen Horowitz' "AI Cannon" is a good reference listicle (https://a16z.com/ai-canon/)
- "12 factor agents" (https://github.com/humanlayer/12-factor-agents)