The ability of artificial intelligence – myth, straszka, or overachievement? On X, I discussed this with Tom Chajek, 1 of the best Polish progs. The sentences were rapidly translated into a debate about what differs from the individual from the algorite: memory, learning and ability to be capable of being able to be active in the early stages. The entourage is amazingly consistent: what we are utilizing today, AI, is the birthplace of the planet – its origin and not its fullness.
This is the main issue: ]]>https://x.com/gps65/status/1972782138861998261]]>
Two worlds: a human student and a maschin
Tomek felt that the next step of AI's improvement should be to learn more in the context of the activity – expertise in long investigation projects. A member, working on something that is easy to use, will be able to learn and change his mind. Model AI is active in 2 phases: learning from the background and then utilizing the acquired knowledge. In the field of use, it no longer has its weights, but it only has its consistency.
I have turned my attention to the fact that there is simply a deficiency of this ability to influence itself on the world's most crucial differences between the existing possible and the possible for human society. What we are utilizing present is simply a sessic memory, which is in the December of the Notatium – not fresh connections in the network.
The memory is not a witness
Tomek requested: ChatGPT would have remembered the erstwhile talks, is it not already the world? I replied: there is not adequate memory, it must be written in this cognition in a form of power. The animals have the power to learn present – they do not have a long-term background and activity.
In this dream, the memory of the context is as if the individual were conducting the notebook and would not remember it. It's not a real world, it's the back of memory.
And what about the smiters or algorithms?
The volume was formed by a liquid liquid fluid: if we meet the bones that usage non-neurovian networks, but, for example, the logs of the first government or the standards of forest, would I consider them to be unwitting? I said, "There's no request for architectures, but there's a stimulus or a stimulus from the sun." If not, they are a determinal machina.
Determinism, non-determinism and the world
The Tomek has rightly recognized that the neuter networks are besides dethermetic, and that there is only 1 part of them. I stressed with my colleagues that the light is inactive gradual. Today's potentials have any kind of plant power – like we think or bacteria – but not a human power. For the transition to the higher levels, there is simply a request for a model that will be able to continuously affect its weights as the power does.
If we consider that all that we are creating is determinal reasoning of black boxes, it is the external non-determinal planet that makes them non-determinal. Impulses from the brain all the time affect this dethermist reasoning black box, which can affect its interior mechanisms. In this way, all dethermic machin becomes non-determinal, if it is possible to determinate the outside planet all the time.
Where's AI?
In practice, this will mark a continuous learning experience in the course of action, an investment in memory with an architecture alternatively than an external focus. AI is already working on it. This is just the tusk that I gave Tomkoki.
Why is that important?
Because it will be a actual investigation partner, not a secondary investigation partner. The certificate is not a single certificate, it is simply a process of integrating in the first place. For this reason, the AGI is simply a long way from the beginning – before the nearness of the technology, and then (i.e. the futurcation can’t be achieved): the entrepreneurship.
And what is truly happening in the AI-resolutions?
This is not a pure explanation – present the largest investigation teams are working on what Tomek and I talked about:
- Meta-learning and training – the ability to work in the course of action, not in the background.
- Continual learning – the strategies that can be utilized to make parlames without «catatrophic for omitting» early knowledge.
- RLHF in the Common – reinforcement of learning not after epidemics, but it is simply a continuous process combined with intelligence.
- Neuromorphic composition – The structures that work more like the synapses in the brain, with the anticipation of a different combination.
- Architectures – the transformation of the memory of the long-term memory that was incorporated in the model itself, the replacement of external donor bases.
Other words: precisely what differs from a book with a note from a man learning. It is the second line of investigation – it will transform AI from a tool into a test subject.
We differed in words, not in essence. He said: "You request expertise", and I said: "You request to learn in the course of action". These are the 2 highlights of this sima future. If AI connects 1 to another, it will cease to be a back tool – it will become a fresh investigation sector.
Grzegorz GPS Swiderski
]]>https://t.me/CanalBlogeraGPS]]>
]]>https://Twitter.com/gps65]]>
