![]() |
Forum Index : Microcontroller and PC projects : Deepseek musings
![]() ![]() |
|||||
Author | Message | ||||
Volhout Guru ![]() Joined: 05/03/2018 Location: NetherlandsPosts: 5050 |
We have had the AI we need to solve these for thousands of years.... ![]() Volhout PicomiteVGA PETSCII ROBOTS |
||||
electricat![]() Senior Member ![]() Joined: 30/11/2020 Location: LithuaniaPosts: 299 |
@Volhout ![]() At some point , funny and scarry idea came to my head. AI in role of highly fanatic priest... from my point of wiev it could be more evil and scarry than "Skynet" scenario My MMBasic 'sand box' |
||||
Mixtel90![]() Guru ![]() Joined: 05/10/2019 Location: United KingdomPosts: 7852 |
Mick Zilog Inside! nascom.info for Nascom & Gemini Preliminary MMBasic docs & my PCB designs |
||||
Volhout Guru ![]() Joined: 05/03/2018 Location: NetherlandsPosts: 5050 |
42 . PicomiteVGA PETSCII ROBOTS |
||||
LeoNicolas![]() Guru ![]() Joined: 07/10/2020 Location: CanadaPosts: 503 |
This excellent video from the channel Computerphile explains the deepseek architecture and why is so cheap to train and run https://youtu.be/gY4Z-9QlZ64?si=5ijf5jUDPW9qMzeC |
||||
Mixtel90![]() Guru ![]() Joined: 05/10/2019 Location: United KingdomPosts: 7852 |
Oh dear..... Big problems in little China? Mind you, our government(s) appear to be not much better. . Edited 2025-01-31 03:47 by Mixtel90 Mick Zilog Inside! nascom.info for Nascom & Gemini Preliminary MMBasic docs & my PCB designs |
||||
zeitfest Guru ![]() Joined: 31/07/2019 Location: AustraliaPosts: 573 |
I remain skeptical, I can see they are a good way of accessing knowledge but not a good source of knowledge itself. There was on old story about an early automatic translator project (guess where). To test it, the developers fed in a phrase, to translate it to Russian, and then retranslated the result back to English. So, the phrase "The spirit is willing but the flesh is weak" came back as "The vodka is good but the meat is rotten". ![]() The point was, languages are often inexact logic. That is compensated for, by using lots and lots of probability based crunching. Recently it has taken off instigated by the increased availability of cheap GPUs which were designed to do lots of the simple arithmetic needed for image creation. Unfortunately crunching on floating point numbers impacts their accuracy. Image creation only needs single precision so the GPUs only provide that...as soon as you need double precision the GPUS chuck it back to the main system, much slower !! I think there is bit of a bandwagon fashion sometimes which leads to uh, inconvenient surprises. When computers became available to researchers, weather simulation forecasting started using computed models. The models were taking too long so the researchers started using (faster) single precision calculation to run preliminary results, with the logic being that after a successful model was found it could then be run in higher precision to fine tune it and give more accuracy. But the two approaches gave wildly different results. That was a surprise as the difference between a single-point number and the equivalent double-point number is very small. One researcher remarked that it was if a butterfly wing flap in the Amazon forest would lead to a storm in the Atlantic. A "Large Language Model" no doubt has much more data and inference available, but can it see all the actions in the Amazon ? Hmmm. Edited 2025-01-31 09:16 by zeitfest |
||||
lizby Guru ![]() Joined: 17/05/2016 Location: United StatesPosts: 3358 |
What is knowledge? It's not just data points, however invaluable data points are in gaining knowledge. But the ability to consider far more data points than any human can is likely to result in novel formulations--new knowledge, as it were. Medication researchers are said to be using AI to suggest possible chemical combinations to be examined for safety and efficacy in treating illnesses. Maybe AI isn't going to come up with ready-made solutions, but it may be able to suggest combinations that researchers have not yet considered. The same goes for many fields. Battery researchers are said to be investigating novel cathode formulations suggested by AI. Magnetism researchers are looking previously unexamined interactions. Plant geneticists are exploring AI-suggested paths for feeding a hungry world. It's early days yet. (of course, there's always the possibility that we are living in late-stage human culture now.) PicoMite, Armmite F4, SensorKits, MMBasic Hardware, Games, etc. on fruitoftheshed |
||||
Mixtel90![]() Guru ![]() Joined: 05/10/2019 Location: United KingdomPosts: 7852 |
This is the point. AI, because of its speed and ability to see what has already been suggested, can search for and find combinations that have not yet been tried. It doesn't always know if those combinations will fit what is required as it has no data on them. Hence you can get a recipe for bratwurst ice-cream as a side dish. Yes, the components of the dish are all edible, but the AI can't taste it or get a review on it. AI is an incredibly fast and adaptable moron and is the epitome of GIGO. Mick Zilog Inside! nascom.info for Nascom & Gemini Preliminary MMBasic docs & my PCB designs |
||||
lizby Guru ![]() Joined: 17/05/2016 Location: United StatesPosts: 3358 |
Of course, the only real solution to "Garbage In, Garbage Out" is not to input any garbage. Since AI is to a large degree fed on what is on the internet, preventing garbage-in is a major problem, especially since in some areas there is considerable dispute over whether some "data point" is or is not garbage. PicoMite, Armmite F4, SensorKits, MMBasic Hardware, Games, etc. on fruitoftheshed |
||||
twofingers![]() Guru ![]() Joined: 02/06/2014 Location: GermanyPosts: 1573 |
I am wondering if there will be commercial AI devices available for around $2000 in the near future ...? Similar to PCs 40 years ago. Michael causality ≠ correlation ≠ coincidence |
||||
DavidCP2010 Newbie ![]() Joined: 09/02/2025 Location: United StatesPosts: 3 |
Yep. This is my stickler with calling "AI" Intelligent. Being able to apply and validate what is known against what isn't is a pretty important part of intellect (arguably the most important part) and yet current AI can't actually do either of those things without a proven model. Some methods (like sim2real) are a little closer I guess because they have an approximated (often too approximated) model to validate against but that falls more into robotics than what people call "AI". A very complex encoding and replay can fool a lot of people (and perhaps be good enough for products) but it's not nearly enough for me to call it smart.😅 |
||||
![]() ![]() |
![]() |
![]() |
The Back Shed's forum code is written, and hosted, in Australia. | © JAQ Software 2025 |