Artificial Intelligence ML

Updated 23-Aug-2026

Artificial Intelligence and/or Machine Learning

See also

• Artificial Intelligence Machine Learning
• Deep Speech
• Tesseract
• Generative AI in Music
• AI Chat Bots
• Local LLMs

The anti-search engine

Put a basic search into Gemini (or any AI chat bot) and compare that with the utter slop of Google or related search results. The quality difference is stark. Yes, there can be errors in AI chat, but there are certainly more of them in search, and any error in the AI side is due to the absolute rubish of search results. I fail to see how AI is sloppier than web search slop.

Google and others have truly destroyed their search brand value through relentless money-grubbing. Frankly we are in a really great time where AI is mostly free and doesn't have ads entering the result stream. Thankfully local LLMs will free us from this situation ever happening, more or less forever (more on local LLMs below).

Context and voice interface

The AI chat bots can understand what you mean, and do what you say, when you say "recheck the information and make sure to include data only from academic and government publications". This can be done using voice or text, desktop or mobile.

In addition, the chats are saved so that one can start on one device and continue on another.

What (else) is AI good for?

Well, it appears, many things.

First off, lets just say using the term AI is really only equivalent to the term machine learning, plus sometimes specifically both small and large language models. We are not dealing with a superintelligence, just a super tool. There is no consciousness here, though it can seemingly simulate it, fooling foolish humans.

That said, we have harnessed massive computing power to generate both helpful and astounding insights. Anyone who is simply dismissive or repulsed by AI will end up looking like luddites who never saw the usefulness or promise of the Internet and the web.

• Generative AI is quite helpful in music, images, and video. Yes, one needs to learn how to manage the prompts effectively, and which models and parameters work well, and it cannot on its own create at the level of human artists. That said it is already being put to use as a very strong assistive technology, and has the promise of drastically reducing the costs of production.
  • Dreams of Violets CGI would have cost millions, it cost $2,000 using AI, this film was made in 6 months, screening at film festivals.
  • ‘I see the incredible promise’: on set of an AI film shoot as new studios embrace controversial tech bypass industry giants and take creative risks.
• Coding assistance is useful in speeding up code generation, and the use of multiple models allows for checking for mistakes by a different model, and even having different agents converse with each other.
• AI can do translation and editing of nonfiction and fiction.
• A huge variety of science and engineering opportunities using AI.

What AI is not good for

• Doing your homework for you - exceptionsally bad at this as it negatively affects test scores, significantly.

Using 30 months of panel data on 26,811 Chinese students in grades 7-12, we study how generative AI affects homework productivity and learning. The data combine monthly closed-book exams, high-school and college entrance exams, and homework scores and completion time across nine subjects. We exploit staggered AI adoption in a difference-in-differences design. AI adoption raises homework scores by 18% and reduces completion time by 30%, but lowers monthly exam scores by 20% within six months. High-stakes entrance-exam scores fall by 18 and 24%, with the full penalty emerging only after about two years. The losses are largest in social science subjects, followed by STEM and languages, and are especially large for junior students, high-achieving students, and boys. The learning losses are concentrated among roughly 80% of AI users whose behavior is consistent with homework outsourcing, as indicated by exceptionally short homework completion time coupled with high homework scores. AI users who maintain similar homework completion time as non-AI users experience small learning losses.

Selection, training, use

The selection of AI models (based on features as well as affordability (generally a hardware constraint), AI training (even after a model is generated), and the use of AI models: all of these are quite labor intensive with a steep learning curve. As someone who wants to install, configure, and understand the system as well as put it to good use this takes a lot of time. However, it is really, really interesting and while I am sure AI overall is a mixed bag of wins and losses, the gains will make permanent changes in how we go about things. There are too many advantages and opportunities, ignorning the hyperscaler and neocloud nonsense, and gold rush psychosis.

Local LLMs

I paid $350 for an Aostar GEM10 with 32gb of integrated ram (no SSD) on October 30, 2025. This is a mini-pc in a small 4"x4"x2" box that sits on my desk. I put a 1gb SSD drive inside and then attached all my peripherals. This device has an AMD Ryzen 7 6800H with 32GB LPDDR5 6400MHz RAM and AMD Radeon 680M integrated graphics (it also has an OCULINK Port if I ever want to spend the money for a dedicated graphics card). My point here is that this is not really even a gaming rig, it is a mini PC, nothing special, for ~$500 USD (yes, prices have gone up since then).

The thing is with shared integrated graphics I can go into the bios and tell it to dedicated 16gb of the 32gb to graphics. I can then offload the entire LLM onto graphics, provided I'm using llama.cpp and having compiled it with VULKAN support. Sure, this is DDR5 speed ram, not the dedicated graphics card DDR6, nor is the GPU very fast, but it can fit the 16B LLM models and run them at a reasonable speed. No need for tears about massive power or water needs for these local efforts. This mini PC can even run two models simultaneously, with a little tweaking of parameters, which means a coding auto-complete and a coding q-and-a simultaneously, as a single use case.

General AI

Artificial Intelligence (AI) was always a joke in the 1990s and 2000s. The main point is that the entire project was a failure. To continue to get grant money, researchers rebranded what they were doing cognitive sciences and learning sciences. The problem at the time was that they could not produce anything resembling intelligence.

In the 2010s that has significantly changed. The main reason for the resurgence of AI is the dramatic increase in computing power and in data availability (largely due to the dramatic increase in sensors and data-creating devices). Also, a more focused aspect of AI is now in vogue, deep learning, which is about learning data representations rather than task performance. In a sense this is much easier to do, with the learning happening at a higher cognitive level, and which can include directed, semi-directed, or non-directed machine learning.

A general purpose learning machine, even one constrained to things like a voice-enabled assistant, has a long way to go for any kind of measure of success. However, certain focused forms of deep learning, namely pattern discovery and anticipation, have had enough successes to generate continued investment. Finance and transportation are two systems which while productive using humans -- the lowest-cost, 150-pound, nonlinear, all-purpose computer system which can be mass-produced by unskilled labor)¹ -- can be improved dramatically using AI applications (neural networks, and computer vision and cybernetics, respectively).

It seems that now is the time to integrate AI, where and when possible, into projects of various kinds. Some basic principles and current state of the art is needed to get one's bearings in AI, and provide a foundation for experimentation with AI in new products and services, or simply useful utilities.

China, LLMs, Robotics

• As with consumer drones, electric cars, and renewable energy, we are dealing with a juggernaut. Qwen especially in LLMs but also we see a lot of progress on robotics as well.
• A fantastic article of someone doing more fundamental AI research (not transformer-based)
• From an interview with DeepSeek's CEO. Indeed, the entire interview is quite eye-opening, though at the same time entirely predictable. In these three markets: drones, EVs, and LLMs, the secret sauce is doing fundamental, architectural research with confidence. That creates disruptive breakthroughs. Of course hiring geniuses who are motivated to do such is mandatory.

large models made in China will be as much of a force to be reckoned with as drones and electric cars.