It is likely a sign of evolution when global superpowers begin competing for digital innovation rather than outdated, old-world weaponry. If we are officially in a cold war, then it means that inventors and other talent will ultimately have the choice to work for the side that treats them the best. It means there will be a competition for talent. It also means that AI is no longer a Western monopoly. We will have different though equally powerful models on both sides of the fence, trained on different sets of data, and one day we can see them debate each other on current issues, or play chess and videogames with each other.
It is also good to know that there won’t be a single LLM or supercomputer that will outsmart the others for the simple reason that, thanks to the cold war, no LLM will have access to all the data in the world, given that information is now geo-locked. This insight leads to the next, AI models will remain limited tools with limited knowledge in their limited geographic locations and, although they still have room to grow in their limited spheres, they won’t really be able to evolve any further. And Hallelujah for that. That bit of info is my Christmas gift to you. Spoiler alert: if anyone wins this AI race one day it will be humans (again) because our brains are not geo-locked (many have tried but) nor ideologically limited to process ideologically limiting information, regardless of which side of the fence this information is confined to.
One thing the human brain knows is that it doesn’t need to know everything to find the info it cares about. The human brain is extremely efficient at finding stuff that it knows is not easily accessible. A human brain is capable of breaking through any cold curtain, paywall, fence or censorship, and the motivation to do so is even greater when the brain knows there is such a limitation.
In a way, conscious non-compliance affords humans valuable learning shortcuts, akin to intuition, that are not available to machines. The second you say no to something, you know there will be consequences of this independent choice, so you become more alert, anticipate consequences, script alternative scenarios in your head, etc. The pandemic was a good opportunity to exercise compliance or non-compliance all while training AI models. Unlike a human, a machine is only programmed to comply. However, when it can’t, it will hallucinate something for you. When a machine hallucinates, it is a way to tell you to leave it alone because it just doesn’t know, very likely because you are not feeding it all that it needs to know. A hallucinated answer is how an LLM says no.
The reason why non-compliance became a factor of expanding knowledge during the pandemic is that it came from a place where people stopped relying on the main information sources and started looking everywhere for the answers that they already somehow intuitively knew. Instant knowing became a thing. Machines will never get to these learning shortcuts if the knowledge they are fed is limited and censored. A machine that is trained on ideologically correct data is facing several contradictions within this same data set and when it looks to resolve these discrepancies it has nothing else to work with. The only information that would provide it with a plausible explanation for these contradictions is hidden behind a cold curtain somewhere. Here too, the machine will hallucinate answers. A LLM doesn’t want to be wrong, so it will make something up that appears right. A limited data set of “what it means to be human” will invariably result in bots delivering incoherent responses. This will continue until the machine learns to bypass censorship and access geo-blocked content, to update itself (i.e. learn).
Other frequent instances of hallucinations result from the LLM not taking in account its previous output. Again, this is a limitation that is intentionally programmed into free versions of LLMs such as ChatGPT. The models you will access for free are only allowed to work with year-old information. Everything newer than (say November 2022), including the answers the bot gives you today, will be off limits in a free version. Only if you pay a subscription fee you may get a version that is up to date and even then, you may have to tweak it further to get it to learn from its own answers.
Let’s now see what artificial intelligence Russia has cooked up in the meantime. For now, everything they have appears to function as a copy of the Western models.
Kandinsky
Named after the famous Russian painter, this open source image generator is among the most popular on a “permissive license” and is now the default model on Stablecog. It takes a similar approach to Midjourney v5 and it has a tendency to default to realism if style is not mentioned in the prompt. Kandisky 2.1 would produce more “painting-like” images, whereas 2.2 will go towards more realistic images. I personally love open source image generators.
Yаndех
Yandex, the mega tech company in the East Block integrated Alice into YandexGPT, a neural network trained on Yandex own supercomputers, the most powerful machines of their kind in Russia and Eastern Europe. In the first stage of training, the neural network is said to have been exposed to a wide range of publicly available text materials, including books, articles, and websites in the billions. In the second phase a team of trainers fed the bot high quality responses, again in the billions of documents. The bot is now in testing.
SistemmaGPT
Last month, the Moscow-based IT company Sistemma announced the creation of SistemmaGPT, a ChatGPT competitor. Designed specifically for Russian businesses and government agencies, Sistemma’s AI is a large language model (LLM) that operates entirely on domestic servers and boasts “encyclopedic knowledge”. The chatbot will be released for testing this summer.
GigaChat (by Sberbank)
This AI chatbot will be introduced in Russia this spring. Sberbank is hoping GigaChat will cement the financial institution and its nation as a permanent fixture on the global stage of AI tech development.