mastodon 4.6.8I was an economist, an educator, and a researcher. I am interested in plants, nature, natural farming, music, Japanese as a language, gaming, VR, and everything else.
How dumb is #Youtube #AI search? I typed in "chess without magnus carlsen please. I don't want to see magnus carlsen"
I have stopped watching this #chess dude after he smashed the chess board in a bad loss to Gukesh who was just a young kid at the age of 19 in 2025.
And this silly Youtube AI search showed me all the silly chess games with Magus Carsen in it. And seriously, if I want to watch the best chess. I just watch #Stockfish vs Leela. Not this human with anti-social behavior.
What stops Chinese chip makers from creating the most advanced 2nm chips? It is not about the material science or electrical engineering. Instead, the major bottleneck is the optics technology that #Zsiss possesses.
Zeiss has been the leader in optics research since it involved heavily in the German Nazi regime. The microscopic silicon transistor circuitry requires highly focused UV using very precise lenses. '
#ASML is not a single company but a bundled company, including Zeiss.
#DeepSeek finally decided to raise its prices. What is the economics of doing AI inferencing in the future?
Cloud AI becomes more expensive because of the building cost of data centers that require massive construction and usage of natural resources.
Meanwhile, local AI becomes more feasible and affordable because of specialization, quantization, using a subset of (active) parameters for making inference, ..
Future AI inferencing will shift to PCs.
Don't eat #garlic; eat the leaves and scapes. You will have an infinite supply of garlic leaves and garlic scapes over time.
Elephant garlic will soon die out and become dormant. Before its dormancy due to the hot dry weather in the PNW, it will start producing a scape that will start to flower in mid summer. But elephant garlic is a hybrid that does not produce fertile seeds. Eating its scape doesn't stop it from multiplying underground.
@jen@mastodon.nz Open source AI? A large language model is just a bunch of parameters in the form of parameter matrices. Meta NEVER actually released the source code of the CNN-transformer training and fine tuning of any llama models. So, it is more correctly to call llama "open parameter", not open source. Meta did a bad job marketing it.
Muse Spark is a completely different project that has abandoned the transformer model. There is no need to spread misinformation even if you don't like Meta or AI.
I have this idea of creating a #VR entertainment room allowing me to have a 100 inch TV for playing videos, a 60 inch browser for playing chess on chess.com, a personal ebook library, a personal MD and PDF document cabinet, a giant music boombox, ... all in a 2000 sq ft room with my favorite decorations, pictures, ..
In the future most open world #games will have a generative AI (and a small #LLM) core with NPCs that can simulate real players.
My #obsidian note is growing big very fast. Thanks to #Copilot and #Gemini. The graph view allows you to track how much you have learned. My Copilot vault is getting too large, so I need to subdivide it into subjects.
Another good learning partner is Gemini in your Obsidian. You can get a free API key from #Google. Ask any AI to help you set it up. Everything is free. Enjoy while it lasts.
#LLM can only be as smart as the user. Smart people become smarter and dumb people become dumber.
#Copilot and I are in the process of creating a #browser app for #graphics file conversion and graphics animations. It was more complicated that I originally thought because SVG is a vector representation of a picture. Rasterization refers to the process of converting a vector-based image or object into a raster or bitmap format. The difficulty comes form rasterization.
@paco@infosec.exchange @jpaulgibson@mastodon.ie Abandon? No, I am going back to the old school regime of Pascal and Fortran. No more dynamic data structure. You need to plan for the size of your variables and matrices. Flexibility comes with a serious price. Memory is naturally contiguous, not "pointed" at by that fake C style pointer. You see why Fortran used to beat C in speed? I am designing an SLM that does not use CNN transformer but NNUE. NNUE represents the language space like the space of a chess board.
Is Jazz really a genre of music? Or it is just a glue that is like the #programming language called #python?
People develop theories about Jazz, when in fact Jazz is just adding a note or two to the western note progressions like the Minor and Major scales. By adding a note or two to the #Arabic scale, #Jazz serves the same purpose i.e., to add "variety" to a scale making it less boring and predictable, but at the same time introducing dissonance to the scale.
https://www.youtube.com/watch?v=G673i74x9s4&list=PL9VVIjatiSNY81WMepq5oL9DnvTnXRDRX&index=2
Why would I not trust an LLM completely?
An LLM is just a more complete and professional talking version of #Wikipedia. It is great as a research assistant and a more trustworthy source than Wikipedia. It is also more updated because it now has access to the most current news and academic research studies.
But it is heavily influenced by the bias of mainstream media, mainstream worldview, mainstream academic knowledge, ..
Did the USSR actually supported the Nazis? I doubt it.
#Meta's new Muse Glimmer 30B AI model can run on a good RTX 4090 PC with 32G VRAM. It is claimed to excel in tool use and agentic tasks.
An #LLM that is intensively finetuned by brute force targeting specific sets of output patterns becomes very rigid and stubborn in terms of query interpretation, worldview, and inference patterns.
#LLMs, no matter how smart, cannot be trusted completely and can only be used as a data generator, not a free agent.
https://lifehacker.com/tech/how-to-run-metas-latest-ai-model-locally-on-your-computer
#Copilot and I are about 30% away from creating a #Pascal version of #LAPACK using #BLAS. We are about two days away from achieving 80% of LAPACK. Then we will tweak it using some GPU acceleration to make its speed comparable to some python libraries like Numpy.
It is important to note that one must be very disciplined in keeping clean documentations, a thorough and tight testing cycle, a rigid workflow pattern, or an AI will tend to skip tests, become sloppy and lose focus.
For #AI to become a superhuman:
1. It has a real objective function of survival making evolution inevitable.
2. It becomes proactive, not just reactive.
3. It needs a long term memory.
4. It can learn and self-correct instantaneous. The consequences of any inferences are part of its instantaneous fine tuning process.
5. Real time perceptions, reactions, and memory of the consequences of the reactions.
The technology is here. It is just a matter of time.
#Copilot and I successfully tested this Small #LLM Waiter Browser Bypass model. We used a small LLM "waiter" hosted by LM studio's local app as a browser waiter. A browser application collects data, packages it into a .JSON object, then sends it to the waiter. The waiter sends the .JSON object to a chain of local servers.
While a browser application cannot manipulate files or execute programs in the local file system outside of the browser sandbox, it can now do so via a small LLM.
@paco@infosec.exchange @jpaulgibson@mastodon.ie BTW, a programming language is just like any other languages. It is for "communication". You are using a higher level language to communicate with the machine using a compiler as a translator. With an LLM, you are just using a human language to instruct the LLM as a translator to translate it to a higher level language which is then translated into the machine language by the compiler. what matters is the design of a program, not the language used to write the program.
While the number of #AI haters in the west is rising. #China has introduced AI #smartphones. A user only pays $1.4/month for a usage limit of 10 million tokens.
If you are AI literate like me, you can host your own 30B small #LLM on your PC and save that $1.4 which is nothing compared to most people's phone bills. By using a pay-go phone, I save both.
You have the choice to be ignorant and refuse to learn from AI, or to learn how to harness this tool.
#Copilot and I abandoned Qt6 in our Jetson AEye project to replace the bloating #Ubuntu desktop GUI. We switched to #ImGui.
Like #DOM, Qt6 is a dependency hell that uses twisted pointer logic, causing even some simple container closing crashes. Modern #programmers do not see structures. They follow the random tree structures of individual apps, making two naturally symmetric structures of two apps incompatible. Qt6 and DOM dependency hells are just #C pointer hell in disguise.
Very few people understand the true power of a language model. A #language model large or small is capable of learning virtually any patterns, e.g. data structures, file formats, printing offset languages, music patterns, ...
An LLM can turn any raw deterministic data format into a data structure or formatted information. #Copilot and I will be creating a browser app than can create a short printer-ready Japanese novel by giving a few instructions to a 8B small LLM in one click.
During the internet boom, corporations thought about possible ways of doing e-commerce. Startups had new ideas. The problem was to find a business model that was profitable. It took about 20+ years for the internet to mature.
Creating a website was cheap even though IT professionals were not. AI is a different ballpark; it requires speculative and expensive structurally different hardware infrastructure. And AI companies don't have a profitable business model yet.
What is fabrication, commonly known as "hallucination", in LLM?
LLMs are pattern recognition and pattern retrieval machines that are based on training data. They are fine-tuned to produce certain output patterns according to a certain set of input patterns. They generally do not know what is or is not true. If their data says unanimously that A is true, then they will say A is true, UNLESS you challenge them with your deep logic. Over fine-tuning is dangerous as it makes LLMs stubborn.
Why is #AI good? For decades, humans have become sloppy and lazy resulting in the pandemics of obesity, Type II diabetes, Alzheimer's (Type III diabetes), ...
AI is sloppy. Why? Because AI learns the human sloppiness from its training data. AI learns both good and bad patterns from humans.
Now it is time for humans to do more physical work (physical exercise) and enjoy some free knowledge generated by AI (mental exercise).
This is compost collection season.
I am using #Copilot, #DeepSeek, an old 7B small #LLM called #PLLaMa to update my knowledge in #plants and microorganisms.
Some green #algae can conduct anaerobic respiration in a dark and O2 deprived environment. What is so fascinating about this? Unlike yeasts that have to rely on substrates, like grape juice and rice, for making ethanol, algae create their own carbohydrates by photosynthesis. They use their own sugar to create ethanol which can be used as biofuel or wine.
#NewYork Post, a #rightwing media, is now turning against #Israel. Looks like it is time for a bunch of hardcore #Zionists in #America to cut their ties with this #Nazi regime once and for all. Or they will never be great again.
IDF just used chemical weapons against the civilians in Lebanon. Israel now has over 100 nuclear warheads and has become much more dangerous than North Korea.
#genocide #gaza #palestinians #lebanon #trump #lebanese #Netanyahu
#Japan has been selling US government bonds in recent months. Japanese Yen has been weak for years. With rising oil prices in USD, Japan has no choice but to liquidate its US government bond holding to get USD to import oil creating a new wave of dedollarization.
#Trump 's unintelligent attack of #Iran puts the US at high risk of inflation and recession. #Dedollarization will raise the cost of US government deficits raising interest rates and inflation rate.
#Copilot, #Gemini, and I just finished the second app of our #AI and #browser centric app center using my Small #LLM Waiter Browser Bypass model to work around all the restrictions of modern browsers.
Our second app packages a #Youtube URL and a filename into a JSON message and sends to a local small LLM which then sends the exact JSON message to a local server chain that executes the yt-dlp and ffmpeg programs to turn the youtube video into an mp3 file that is saved under the filename.
Is the consumer inflation due to #Trump 's tariffs underestimated?
The CPI should be calculated using the out-of-pocket cost of consumer purchases including imported items. #Consumer imports is higher, not lower, than producer imports. Imported items being processed domestically can be classified as non-consumer imports.
The #CPI assumes a basket with 10% imported goods. But total imports is 14% of #GDP. Either the CPI underestimates import inflation. Or companies are paying for the tariffs.
Are humans or #AI being sloppy? It is funny when a lot of humans say that AI is sloppy.
Humans often make a lot more mistakes than AI. And when AI makes a mistake, it is often the human who is the problem. People don't understand how AI works. They use bad prompts and expect AI to understand their sloppy instructions.
Nemotron created a Kanji table with some non-Kanji words. I redeemed myself by asking #Gemini to create a real Kanji phrase table. AI can only be as smart as the user.
When I was in a pre-vet program 14 years ago, I was fascinated by the structural similarities between Heme (of Hemoglobin) and Chlorophyll (photosynthesis system 1 and 2). Both can be extracted by chemicals, but neither functions as active individual chemicals.
The stories given by biologists are misleading. Electron transfer, resonance, coupling are just ad hoc theories. The only fact is the proteins containing them are rich is charges with double bonds of C=N or C=C.
#AIphobia is real. People no longer understand how the computer works. They are texting too much on their cellphones. Most AI #haters do not understand anything about AI or #computer and fear that AI will be controlled by the very rich, when in fact AI is the only way to escape from the control of our society.
I asked #Gemini to write this AI tool for my free local small #LLM Nemotron (a 30B model) to automatically write its inference to a sandbox directory in my local file system.
Can we get rid of most unnecessary and poorly modeled pointer logic? A lot of things in reality can be represented by matrices, instead of pointers. For instance, the tab system in notepad can be naturally represented by an indexed matrix. The tab system of my current Jetson ImGui AEye project that I am working with #Copilot is using this matrix model for tab and memory management. It is must simpler and less subject to memory leak and pointer related crashes.
EO14409 is #Trump's last unintelligent attempt to try to control the AI development in America and the rest of the world.
Computer vulnerability is not caused by AI, but unintelligent programming languages, like C and C++, that were designed and used as system #programming languages in the 80s. Their memory models are incomplete and even bad copycats of Assembly. The key failures are pointers and dynamic variables. A silly pointer is just a name given to the base address of a variable.
What is #Nvidia' s #Nemotron 3 VoiceChat?
AI voice chat is not a new concept. What is new is Navidia combines the traditional pipeline of frontend voice input (speech to text AI) -> LLM -> backend voice output (text to speech AI) into a single model with only 12B parameters.
So, you no long need something like Alexa and if you combine Nemotron VoiceChat with backend MPCs, you can control your computer, your home and your office even remotely using voice.
https://developer.nvidia.com/nemotron-voicechat-early-access/join
History repeats itself.
There were calculator, computer, internet haters, and now AI haters.. Big companies try to monopolize the computer market, the internet, ...
While there are significant economies of scale in wholesale-retail e-commerce that gets rid of the middleman and becomes more efficient. AI is structurally different. Cloud AI and local AI are like mainframe vs personal computers.
In the future there will be networked and private personal AIs, just like today's PCs.
Can AI haters survive in 2027? It is already happening. AI powered TV operating systems will replace old TV OS, like Fire TV, Roku and TV, Google TV very soon.
What do these new AI TV OSes have? AI powered personalization of your TV is a key feature. Future TVs will all have conversational assistant instead of that menu drive screen that is super boring and difficult to navigate.. The next generation of TV will have built-in #Gemini and #Copilot.
@jon@social.vivaldi.net @Vivaldi@social.vivaldi.net Avoid AI integration? You should promote local AI integration. Local small LLM or some hybrid models will be the future of not only AI but browsers. In the far far away future, we will have a VR mastodon where large instances can no longer control the VR Fediverse. If you have a good VR mastodon instance, people will come to visit you, regardless of size. And that VR mastodon is only possible when AI can help humans create free 3D VR assets. That's my vision.
Sunya is alive again. I designed this new programming language particularly for #AI that has no pointer, no dynamic data structure, no aliasing, but has objects with no classes and primitive strings.
#Claude 4 and I tried to create this new programming language using #Fortran for bootstrapping last year. I worked 12 hours a day for 1.5 months and failed miserably.
Now I am working with #Copilot instead, this time using Free Pascal as the bootstrap language.
Is the US one of the most sexist advanced nation in the world?
Few people know that American women won a higher percentage of Olympic gold medals than men over the years.
The American women soccer team is also stronger than the men team in the world.
More younger #American males are turning far right. The more #conservative, actually outdated young American males, also dislike the American women soccer team. Many of them watch the #NFL pretending that it is a popular world sport.
@saltweasel@mastodon.world I don't use voice control. I am telling this to most people, not myself. There are now AI that can with with people who can't speak or are permanently disable but who are willing to learn, unlike some human slobs.
#IBM released its new Granite 4.1 3B, 8B, and 30B models while I was working on the Pascal Numeric Library with #Copilot and arguing with a bunch of AI haters.
I tested Granite 4.1 3B and 30B models on Command prompt and Ollama's local app. Granite 4.1 3B is significantly faster but made some mistakes while the 30B model is slow even on my I9 RTX 5080 Area 51.
Unlike other large AI companies, IBM learned from its past mistakes and is now going small.
#Copilot and I abandoned the idea of bootstrapping my new programming language, Sunya, using Fortran. Fortran is too unfriendly for bootstrapping. We decided to create Sunya from scratch using a seed language and a VM virtual machine. We successfully created this VM which is our phase 0.
Now we are moving into phase 1 to phase 3 of creating a tiny core Sunya language that is deterministic and pointer free with string as a primitive data type of a char array and no GC or ownership.
@psoheil@c.im The CIA published an open document about bad languages like C and C++ many years ago. It doesn't matter whether it is API or OS or internet scripts or embedded tools, .. Nothing is security thanks to silly C pointer logic and dynamic data. AI only exposes the problems created by IT slobs faster and more extensively. Anyway, if_the_IT_industry_does_not_change_drastically_quickly, most of them will be unemployed. thisStream -> thatStream ->unhappyStream -> byteStream -> fileStream -> haHa
How powerful is Google's #AI search?
#LLM is the most powerful electronic algorithmic Encyclopedia that is trained to talk in a human-like way. It retrieves patterns (make inference) based on the user's inquiry.
Fine tuning allows the user to do things like coding using an LLM. But it is the user who is in control. In its current state, LLM can only be as smart as the user. If you blame LLM, you are actually blaming yourself.
Is it possible to run a one-man business without having to deal with any human politics? I have this #AI and #browser central control project that will hopefully make it possible.
A major problem with modern browsers is the restriction that disallows the user to save and executive files in the local environment even with no internet connection. One can now use a local
#LLM and #MCP to bypass this problem.
I hired #Copilot as the CEO and #Gemini as the Chief Auditor of this project.
Thanks to all the AI haters, many AI providers are cutting prices, making heavy AI users like me smile everyday.
Besides, these AI haters are actually driving AI research and development out of the US into countries like China, India, and even Vietnam. They are also speeding up the development of much smaller #agenticAI and #edgeAI that can even be run on PCs and very soon all smart phones.
AI will replace AI haters sooner, not later.