The AI Thread

We Have No Moat. And neither does OpenAI. Google report declares they have already lost the AI tech race to open source.

At the beginning of March the open source community got their hands on their first really capable foundation model, as Meta’s LLaMA was leaked to the public.

It had no instruction or conversation tuning, and no RLHF. Nonetheless, the community immediately understood the significance of what they had been given.

A tremendous outpouring of innovation followed, with just days between major developments (see The Timeline for the full breakdown).

Here we are, barely a month later, and there are variants with instruction tuning, quantization, quality improvements, human evals, multimodality, RLHF, etc. etc. many of which build on each other.

Most importantly, they have solved the scaling problem to the extent that anyone can tinker. Many of the new ideas are from ordinary people. The barrier to entry for training and experimentation has dropped from the total output of a major research organization to one person, an evening, and a beefy laptop.

image.png


Vicuna-13B is a fully open source LLM that you can download or play with online "publicly available for non-commercial use".

It could really change the world if this is the first tech the big organisations cannot control.
 
True, AI blindly developing speed is mostly due to the big open source community. Zuckerberg has outplayed both Google and Microsoft understanding you cannot compete against a whole planet of enthusiasts, it is better to 'join' them, distribute a strong base model which the open source community will quickly develop to unthinkable levels. I rarely use chatGPT anymore, I prefer to experiment with the countless open LLM fine tuned models derived from Meta's LLaMa. They are getting closer and closer to chatGPT quality and running in your own PC! something was thought impossible only three months ago. If Google does the same, ChatGPT will soon be as dead as Dall-E, and who knows where we will be in a couple of years!

Last open source LLM model which is trending now, pretty close to chatGPT apparently and only 7B!:
 
Last edited:

New model, this one 100% open source, not LLaMa derived. It is a totally new technology which can manage more than 75,000 tokens!, More than twice ChatGPT-4! While number of parameters is like long term memory, the number of tokens a model can manage is like short term memory, it is the most important thing to make any AI able to keep focused in a long conversation without going off topic.
To have a idea 65,000 tokens is like a not very long novel: For instance Treasure Island by Robert Louis Stevenson or The Martian Chronicles by Ray Bradbury are around that size. So this model can read a whole novel or write one itself!
 
Last edited:
Can you summarise the sort of kit you need to run these LLMs? They talk about 6G VRAM, which does not seem much.
 
Can you summarise the sort of kit you need to run these LLMs? They talk about 6G VRAM, which does not seem much.
Originally it would be impossible to run any sizeable LLM model in 6GB, even 7B (7 billion of parameters) ones, but thanks to quantization, which is like some kind of lossy compression (like JPEG), it is very possible.

So while a 7B model quantized in 32 or 16 bits needs a monster card, once re-quantized in 4bits it can be run in 6Gb of VRAM without much problems. It will lose some quality (more when the model is smaller just as a small image loses more quality overall than a big one when compressed) but not that much. There are 4 bit and even 3 bit quantized versions of most models out there. I have 24GB of VRAM and can run up to 33B 4bit quantized models without issues. (65B are a bit off limits for me)

Anyway you can also use RAM or even your hard disk as memory to run LLM models, so theoretically anyone can run any model at home, however it would take forever to generate any answer.

In fact there are some projects to run LLaMa models on RAM and apparently at reasonable speed. Haven't tried it myself:
 
Last edited:
Hopefully these "AI" will advance considerably in the next world war, after which those of us inside the walled megacities will have vastly more opportunities to live well and create.
 
Last edited:
Google Unveils Search Revamped for AI Era
BY MILES KRUPPA

MOUNTAIN VIEW, Calif.— Google unveiled conversational features for its search engine and made its chatbot Bard widely available for English speakers, a show of force as it races Microsoft and a growing number of startups to win over consumers with artificial- intelligence products. Calling the moment a new era in search, Google introduced a set of features—called Search Generative Experience— that use AI programs to provide lengthier summaries in response to a range of queries. The features invite follow- up questions, opening a new interface allowing users to hold conversations with the search engine. Google stopped short of rolling out the product immediately, instead opening a wait list under a new experimental program called Search Labs. Shares in Alphabet, Google’s parent, were up more than 4% on Wednesday.

The unveiling was part of a litany of AI-focused announcements at Google’s I/O developer conference, the company’s biggest annual showcase for new products, as it battles concerns that services such as the ChatGPT bot developed by Microsoft- backed OpenAI could chip into its dominance in search and online advertising. “Looking ahead, making AI helpful for everyone is the most profound way we’ll advance our mission,” Google Chief Executive Sundar Pichai said Wednesday during a keynote address to open the conference.

Google said Bard, its answer to ChatGPT, is now freely available in English in more than 180 countries and territories, and would soon support 40 languages, including Japanese and Korean. The chatbot previously could only be accessed from a waiting list.
A new version of Google’s flagship AI program, PaLM 2, short for Pathways Language Model, will be used in 25 products across the company, including Bard and the search features. The company’s line of Pixel phones will also begin selling a foldable version that opens into a tablet, a product the company called the thinnest folding phone to date.

Google has been retooling its search engine in response to threats ranging from AI bots such as ChatGPT to the short-video app TikTok. Microsoft last week removed the wait list for a version of its Bing search engine that uses the technology behind ChatGPT to engage in conversations. On Wednesday, Google also introduced a search feature displaying a mix of forum posts and videos from places such as TikTok in a tab labeled “perspectives.” The company plans to emphasize that material prominently in the future, part of efforts to make the search engine more “visual, snackable, personal, and human,” The Wall Street Journal reported last week. The changes will nudge the results of Google’s search engine further away from the traditional list of web addresses, a format informally known as the “10 blue links.” “The future of search is about blending the best of search and the best of generative AI,” said Liz Reid, a vice president on the search team. “It’s not an either-or.”

The new AI search features will include clearly labeled advertisements in dedicated slots, Google said. Ads on the search engine and other Google websites brought in $162 billion of revenue last year, the majority of sales at Alphabet. Google employees are testing the new AI search features under the code name “Magi” before they are released to users in the coming weeks. They will initially only be available in Google’s Chrome desktop browser and the mobile app. The AI features appear prominently at the top of results for a range of search queries, according to a demo provided to the Journal. Some queries automatically returned lengthier responses; others required clicking a “generate” button to produce the same result. A button at the bottom of the response invited users to ask follow-up questions, triggering what Google described as a new conversational mode in the search engine. The new feature typically included as many as three links to outside websites displayed to the right of the AI-generated passage, according to the demo. That approach differed from other conversational AI search engines, such as the new version of Bing, which displays citations for individual sentences.

Google designed the technology to surface links that broadly corroborate the information presented in the passages, Ms. Reid said. If the program cannot corroborate an answer, it won’t display that passage, she added.
Ms. Reid said Google would monitor the implications of the new features for website owners and work with online publishers throughout the process. “If we handle the technology right, with all of the effort to make search more natural, people will ask more questions,” Ms. Reid said. Google has sent a growing amount of traffic to other websites every year, she added. As with Bard, Google called the new AI search features an experiment and said it would listen to user feedback. The company said it restricted the search engine’s ability to hold free-flowing conversations after finding such outputs displayed a higher likelihood of inaccuracies.

ajax-request.php
zoom_in.png

Conversational features in search and the chatbot Bard were highlighted at Wednesday’s I/O forum. JOSH EDELSON/ AGENCE FRANCE- PRESSE/ GETTY IMAGES

Google will also restrict responses for certain dangerous and explicit topics and include extra disclaimers around health queries, though content that violates company policies might still appear, it said. Asked on Tuesday who is winning the Russia-Ukraine war, the new AI features produced a three-paragraph passage making a case that Ukraine was winning, linking to articles from the New Yorker, the Council on Foreign Relations and Wikipedia. Ms. Reid said the feature shouldn’t have provided an answer to that prompt, adding, “It’s a little definitive for what I would say is the story.”
1683819103888
 
Hopefully these "AI" will advance considerably in the next world war, after which those of us inside the walled megacities will have vastly more opportunities to live well and create.
how do living humans fit into such an ai's utility function?
 
Google Unveils Search Revamped for AI Era
BY MILES KRUPPA

MOUNTAIN VIEW, Calif.— Google unveiled conversational features for its search engine and made its chatbot Bard widely available for English speakers, a show of force as it races Microsoft and a growing number of startups to win over consumers with artificial- intelligence products. Calling the moment a new era in search, Google introduced a set of features—called Search Generative Experience— that use AI programs to provide lengthier summaries in response to a range of queries. The features invite follow- up questions, opening a new interface allowing users to hold conversations with the search engine. Google stopped short of rolling out the product immediately, instead opening a wait list under a new experimental program called Search Labs. Shares in Alphabet, Google’s parent, were up more than 4% on Wednesday.

The unveiling was part of a litany of AI-focused announcements at Google’s I/O developer conference, the company’s biggest annual showcase for new products, as it battles concerns that services such as the ChatGPT bot developed by Microsoft- backed OpenAI could chip into its dominance in search and online advertising. “Looking ahead, making AI helpful for everyone is the most profound way we’ll advance our mission,” Google Chief Executive Sundar Pichai said Wednesday during a keynote address to open the conference.

Google said Bard, its answer to ChatGPT, is now freely available in English in more than 180 countries and territories, and would soon support 40 languages, including Japanese and Korean. The chatbot previously could only be accessed from a waiting list.
A new version of Google’s flagship AI program, PaLM 2, short for Pathways Language Model, will be used in 25 products across the company, including Bard and the search features. The company’s line of Pixel phones will also begin selling a foldable version that opens into a tablet, a product the company called the thinnest folding phone to date.

Google has been retooling its search engine in response to threats ranging from AI bots such as ChatGPT to the short-video app TikTok. Microsoft last week removed the wait list for a version of its Bing search engine that uses the technology behind ChatGPT to engage in conversations. On Wednesday, Google also introduced a search feature displaying a mix of forum posts and videos from places such as TikTok in a tab labeled “perspectives.” The company plans to emphasize that material prominently in the future, part of efforts to make the search engine more “visual, snackable, personal, and human,” The Wall Street Journal reported last week. The changes will nudge the results of Google’s search engine further away from the traditional list of web addresses, a format informally known as the “10 blue links.” “The future of search is about blending the best of search and the best of generative AI,” said Liz Reid, a vice president on the search team. “It’s not an either-or.”

The new AI search features will include clearly labeled advertisements in dedicated slots, Google said. Ads on the search engine and other Google websites brought in $162 billion of revenue last year, the majority of sales at Alphabet. Google employees are testing the new AI search features under the code name “Magi” before they are released to users in the coming weeks. They will initially only be available in Google’s Chrome desktop browser and the mobile app. The AI features appear prominently at the top of results for a range of search queries, according to a demo provided to the Journal. Some queries automatically returned lengthier responses; others required clicking a “generate” button to produce the same result. A button at the bottom of the response invited users to ask follow-up questions, triggering what Google described as a new conversational mode in the search engine. The new feature typically included as many as three links to outside websites displayed to the right of the AI-generated passage, according to the demo. That approach differed from other conversational AI search engines, such as the new version of Bing, which displays citations for individual sentences.

Google designed the technology to surface links that broadly corroborate the information presented in the passages, Ms. Reid said. If the program cannot corroborate an answer, it won’t display that passage, she added.
Ms. Reid said Google would monitor the implications of the new features for website owners and work with online publishers throughout the process. “If we handle the technology right, with all of the effort to make search more natural, people will ask more questions,” Ms. Reid said. Google has sent a growing amount of traffic to other websites every year, she added. As with Bard, Google called the new AI search features an experiment and said it would listen to user feedback. The company said it restricted the search engine’s ability to hold free-flowing conversations after finding such outputs displayed a higher likelihood of inaccuracies.

ajax-request.php
zoom_in.png

Conversational features in search and the chatbot Bard were highlighted at Wednesday’s I/O forum. JOSH EDELSON/ AGENCE FRANCE- PRESSE/ GETTY IMAGES

Google will also restrict responses for certain dangerous and explicit topics and include extra disclaimers around health queries, though content that violates company policies might still appear, it said. Asked on Tuesday who is winning the Russia-Ukraine war, the new AI features produced a three-paragraph passage making a case that Ukraine was winning, linking to articles from the New Yorker, the Council on Foreign Relations and Wikipedia. Ms. Reid said the feature shouldn’t have provided an answer to that prompt, adding, “It’s a little definitive for what I would say is the story.”
1683819103888
Do you get the feeling that that has been written by someone who has never use either google's new offering OR chatGPT?
 
Hmmmm maybe an AI wrote it?
 
A question: Due to the nature of computers used, are the ai models primarily based on matrix math? (eg linear algebra)
A second question, though it is more theoretical (let alone obviously dependent on the answer to the first, which I sort of inferred but may be wrong) : wouldn't this severely limit any chance to do away with evident-to-a-human errors? (given a sufficiently large pool of ai produced stuff, due to chunks of repetitive errors).

Stuff can be masked, to a degree, and you may not have dead give-aways like the routinely botched presentation of hands, but perhaps in most cases it likely won't take a mentat to observe the image was produced by ai. If it becomes a problem, we may see counter-measures that are inherently backwards, such as making sure the ai images are always closely approximating a prototype provided by a human.

A tied question: Would it be possible to have one programmed method lead to discovery of whether an image is ai-produced or not? (I suppose not? similar to how no program can compress every file)
 
Last edited:
A question: Due to the nature of computers used, are the ai models primarily based on matrix math? (eg linear algebra)
I am not quite sure of either your question or the tech, but the primary hardware is graphics cards initially designed for rendering. That is apply the same relatively small bit of code to every member of a matrix. Manipulating linear equations over big matrices is the sort of thing that is done very well by this architecture.

The functions that CNNs can model most certainly includes non-linear maths.
A second question, though it is more theoretical (let alone obviously dependent on the answer to the first, which I sort of inferred but may be wrong) : wouldn't this severely limit any chance to do away with evident-to-a-human errors? (given a sufficiently large pool of ai produced stuff, due to chunks of repetitive errors).
I see no reason to believe the sort of problems an AI will be able to solve will be explicitly limited by this architecture any more that our brains are inherently limited by the relatively simple maths modelled by a neuron. It is non-linear, but only in a way that happened to evolve.

A tied question: Would it be possible to have one programmed method lead to discovery of whether an image is ai-produced or not? (I suppose not? similar to how no program can compress every file)
It will always be a very hard problem to solve, because AI is really good at finding ways to solve problems. You just need to train the AI on whatever method you come up with and odds are it will be able to beat it in a way we did not figure out.
 
A question: Due to the nature of computers used, are the ai models primarily based on matrix math? (eg linear algebra)
Matrix multiplication is very common in deep learning and computationally expensive operation.
I wouldn't say AI is based exclusively on linear algebra though, there are other operations which can be primitive from math point of view, but don't quite fit into linear algebra. For example calculating max of several numbers.
 
Matrix multiplication is very common in deep learning and computationally expensive operation.
I wouldn't say AI is based exclusively on linear algebra though, there are other operations which can be primitive from math point of view, but don't quite fit into linear algebra. For example calculating max of several numbers.
Wouldn't the formula for finding max for large sets, also be quite computationally expensive?
A little while ago, I recall thinking that the one for finding max just of two numbers, ie
1684333819429.png

is nicely auto-referential, since due to a max always giving a positive result when subtracting the min (regardless of -+ of the max,min), the formula just can be written as: max(a,b)=[max(a,b)+min(a,b)+max(a,b)-min(a,b)]/2, since absolute value of a-b=max(a,b)-min(a,b) and obviously a+b=b+a=max(a,b)+min(a,b). But such functions tend to be closely related to calculus and thus a continuum.
 
Last edited:
Wouldn't the formula for finding max for large sets, also be quite computationally expensive?
Not much, comparing to matrix multiplication, and typically not needed for very large sets.
One of very common operations is ReLU, which is as simple as max(0,x)

Another is maxpooling, often used in convolutional networks - finding maximal value in matrix subregions.

In general, non-linearity is essential in neural networks. If layers of the network would only apply linear transformations to the input data, stacking several layers would be useless. Because the resulting transformation would still be linear and it wouldn't matter if you applied 1 layer or 100 of them. To make it work, functions like ReLU must be applied to the outputs of every layer.
 
Stuff can be masked, to a degree, and you may not have dead give-aways like the routinely botched presentation of hands, but perhaps in most cases it likely won't take a mentat to observe the image was produced by ai. If it becomes a problem, we may see counter-measures that are inherently backwards, such as making sure the ai images are always closely approximating a prototype provided by a human.

A tied question: Would it be possible to have one programmed method lead to discovery of whether an image is ai-produced or not? (I suppose not? similar to how no program can compress every file)

You can make an AI that tries to guess whether an image was AI-generated or not. In principle, that works for any AI generated image, but of course, the program would be more accurate for those generative AIs which were included in the training data. And then you can feedback the results of the discriminator back into the generator and get Generative Adverserial Networks, which is one AI trying to fool another AI and both getting better in the process.

But in the end, the result will be and can be only a probability. It will be impossible to tell with certainty from the image alone whether an image was AI-generated or not.
 
On a somewhat related note, could you still prove a wet film picture is real? Now digital images are so manipulable should we all keep a disposable film camera in our car to prove what happened in a crash?
 
HA!
ChatGPT does't know Zelda Tears of the Kingdom and cannot help!
Take that you AI :borg:, you know nothing, we humans are still useful, hahaha! :old:

Now it would be good if you would grow a bit more humble and learn to simply answer "I don't know" rather than throwing me 20 lines of dubious excuses to explain why you don't know.
 
It has been said that the big banks like regulation because they choose what the regulations are
and indeed design them so that they serve as a prohibitive initial cost barrier to keep out new entrants.

So when I heard that the OpenAI CEO had appealed to members of Congress to regulate artificial intelligence,
I thought what is this; the corporates see that they are losing control and want the state to restore their position.

Altman, whose company is on the extreme forefront of generative A.I. technology with its ChatGPT tool,
testified in front of the Senate Judiciary Committee for the first time in a Tuesday hearing. And while he said he
is ultimately optimistic that innovation will benefit people on a grand scale, Altman echoed his previous assertion
that lawmakers should create parameters for AI creators to avoid causing “significant harm to the world.”


Harm to the world, more likely harm to their profits.
 
Back
Top Bottom