Not many people realize that the job of an AI engineer is *not* about training AIs. It used to be, yes. Just like once upon a time the job of an AI engineer was about coding heuristic rules, and before that it was just plain logic programming.

We have for some years now switched from curating training data and training #AIs on those to designing continuous processes and systems which allow the AIs to train themselves.

There are many ways to do this:
- Design families of games where AIs can compete against each others. The age of single games has been over for a while now, now it's about making AIs create masses of games each with novel rules and play those.
- Make AIs themselves curate training data from suitable sources.
- Make AIs refine the training data by evaluating it continuously, or by simply making progressive improvement steps on it.
- Make AIs perform easier inverse tasks, and then turn those around into harder tasks.
- Make AIs self-improve both in designing new kinds of tasks for example in domains like mathematics and software engineering, and also in the tasks themselves and in evaluating the quality of the solutions.

In a way, the AIs are pulling themselves out from unreality into the material world, and we're just supporting them in that.

This is the latest in a really good series of essays that consider #AIs, what they are, how they work, and how *we* perceive them. Well worth reading!

https://open.substack.com/pub/tomstafford/p/yes-and-mode-understanding-language

A new technique for LLMs has just landed: Explainable training!

Let me *explain*.

Normal supervised training works so that you show ground truth inputs and outputs to a model and then you backpropagate the error to the model weights. All in this is opaque black box. If you train with data which contains for example personally identifiable information (PII) or copyrighted contents, those will plausibly be stored verbatim in model weights.

What if we do it like this instead:

Let's write initial instructions to an LLM for generating synthetic data which resembles real data.

Then we go to the real data, and one by one show an LLM an example of a real data, and an example of the synthetic data, and the instructions used to generate the synthetic data. Then we ask it to iteratively refine those instructions to make the synthetic data resemble real data more, in the features and characteristics which matter.

You can also add reasoning parts, and instructions for not putting PII as such into the synthetic data generation instructions.

This is just like supervised learning but explainable! You'll get a document as a result which has refined instructions on how to generate better synthetic data, informed by real data, but now it's human readable and explainable!

You can easily verify that this relatively small document doesn't contain for example PII and you can use it to generate any volume of synthetic training data while guaranteeing that critical protected details in the real data do not leak into the trained model!

This is the next level of privacy protection for training AIs!

AIs are not going to save the world, or change it for the better. They are only there because they will work for #billionaires more cheaply than humans do. Everything today is *FOR PROFIT*.

#AIs will not save us. We need to do that ourselves.

https://aeon.co/essays/what-godels-incompleteness-theorems-say-about-ai-morality

From https://t.me/pravdaGerashchenko_en

For the first time, a #Russian warship openly escorted tankers from the '#shadow_fleet' through the #English_Channel, #Mark_Douglas of Starboard Maritime Intelligence reported

He explained that #AIS data revealed coordinated movements between two tankers and a #Russian warship starting June 16, indicating "deliberate timing to allow all three vessels to transit simultaneously, en route to load oil in #Russia."

#AIS data shows that as of five minutes ago, tanker traffic is still moving in both directions within the #StraitOfHormuz. #OOTT #tankers

Hat hier zufällig jemand Ahnung von #Seefunk in Verbindung mit #AIS?

Freifunk-Standort-Neubau: Letztes Wochenende haben viele Hände in #Meinzerhagen einen Richtfunk-Standort nach 10 Jahren Dauerbetrieb um ca. 100m verlegt. Dafür Strom- und Netzwerk verbuddelt. Technik ist in einen alten KVZ59 eingezogen. Station feeded zu https://sdrmap.org #ais #adsb #sdr #freifunk

@capnthommo I've now got my email (paid-for) safely encrypted on servers in Germany behind the #GDPR.

Trump's corrupt tech bro buddies can no longer use my email, etc. to train their shitty #AIs.

Training #AIs, whether #LLMs or #LMMs, is all about data.

They perform tasks within the distribution of their training data, and in the same level of competence.

This also applies to reinforcement learning; if you look behind the scenes, what happens is supervised learning with extra steps. The error is still backpropagated back through the network, but the objectives are slightly different. Instead of direct task performances to imitate, we'll get the feedback from exploration and rewards, basically the most successful explorations can be understood as the data.

So, the data is where the buck stops when an #AI makes a misjudgement. When designing data pipelines, refinement processes and exploratory games for your AIs, you can forget about everything you know about the mechanistic aspects of Transformer architectures, and just simply focus on how to improve the data.

The data is better if it can be used to train better models. Better models have higher competence in skills, higher volume of relevant knowledge, and higher coverage in task generalism and flexibility. Hence the best data represents exactly those aspects maximally.

Human level is not the gold standard and neither is raw real-world data. We can design processes which refine the data without apparent limits, and make that data alive through AIs trained with it.

Obviously, the only answer to #DeepSeek releasing an #AI is to invest even more into creating #AIs. Thanks #EricSchmidt!

https://techcrunch.com/2025/01/28/eric-schmidt-says-deepseek-marks-a-turning-point-for-the-global-ai-race/

It is all about reasoning now. #LLM chatbots are already superhuman in their vast knowledge, and no human can ever compete with their billions of parameters in capacity.

Indeed, we aren't benchmarking their knowledge against single human level, but against the totality of humankind.

But to get the most value out of this knowledge, these #AIs need to be able to execute cognitive skills as instructed on this knowledge. Reasoning is one of such skills, "system 2" in Kahneman's categorization. But it's not just about building an architecture which is in principle capable of system 2 like thinking.

Reasoning is not a single skill, it is an open bag of skills which require mental exploration and which allow producing more knowledge out of already crystallized knowledge.

For example, there are skills needed for solving mathematical equations, multiplication, division and so on. These are different mental skills than skills needed for constructing and evaluating mathematical proofs. Furthermore, reasoning skills related to engineering tasks, scientific work, medical evaluations, and so on, are all slightly different.

Although there are some ridiculously trasferrable reasoning skills which apply across many domains, many domains also have their own reasoning skills which differ from the same in other domains.

We can teach these skills to the AIs similarly as we taught knowledge to them; first by presenting them to them in volumes, and then letting them practice them and improve while doing it.

The same as with humans, but in the end we get AIs which rival the whole of humanity in volumes of available knowledge, but also in cognitive skills, and are able to use their universalist expertise across all domains.

This alone, even without surpassing humanity's intelligence, will alone mean a golden age never seen before. Imagine the repercussions.

Die ADS-B/AIS/Radiosonden-Station im #Wirhaus wurde renoviert. Damit jetzt mit gutem LNA + Weiche und zwei Sticks für Wettersonden, feeding zu https://sdrmap.org #hackspace #neanderfunk #adsb #ais

#ADS-B #AIS #radiosonde release party livestream 19/12/24 @ 1900 CET

We would like to invite you all to join our launch webstream where we will present the new features of version 4.0

We are keen on your opinion and ideas so do not hesitate to join the mumble discussion afterwards.

The stream starts Thursday 19 December 2024 19:00 CET on https://chaos-consulting.de/stream.php
You can join us on our mumble server on mumble://mumble.chaos-consulting.de

AI / LLMs are killing our world, please don't use them.

"...global (data centre) greenhouse emissions between now and the end of the decade are likely to be three times higher than if generative AI had not been developed."

https://www.theregister.com/2024/10/14/ai_datacenters_coal/