Weekly Updates

The latest developments in AI technology have launched the world into a frenzy. From regulators working on bad regulation and researchers testing how good a text generator is in solving math problems, to people, in general, trying to figure out whether an AI overlord will replace their jobs.

Apart from the fear-mongering by some executives we discussed in a past issue, a big part of this chaos arrives from not understanding the underlying technology and what is possible.

So today, we’ll take a look at five issues that are a source of miscomprehension about AI.

Autonomous AI

The idea of an autonomous, Skynet-like AI deciding to destroy humanity is really catchy but should be left for movie scripts since it completely misses reality. Even though there is space for semi-autonomous decision-making bots like the Perseverance Mars Rover, at the end of the day, we’ll always have a human defining what that AI needs to do because that is why the AI was built in the first place.

The term artificial intelligence makes us think that an AI is a semi-human because that is our reference for intelligence. But at the end of the day, contrary to the original Mechanical Turk chess player, an AI is nothing more than a massive table of probabilities. And a table of probabilities does not have a will of itself. The will to do something is always going to be the will of the operator of that AI. In the case of ChatGPT, that will is of the person sending a prompt to it.

AI, like a knife or any other technology, is a tool to amplify what a human can do. We could even say it is a tool to amplify humans. And in the same way that that knife does not replace the chef, AIs won’t replace humans.

In a previous newsletter, we said:

“A safer way to use such tools is in a supervised manner. The user asks the model to generate a start point automatically and then reviews (supervises) what was produced, verifying and ensuring it does what the user wants and not what was most likely probabilistically. […] By itself, a model can degenerate into probabilistic garbage with real-life consequences. With supervision, it can be an exceptionally powerful asset.”

That was a really technical way to say that AI is a tool used to achieve a purpose defined by its operator.

Another helpful example of using technology as human capability amplification comes from airplanes. Airplanes are some of the most regulated and safety-checked pieces of technology that exist today. Similar scrutiny is only reserved for things like nuclear reactors. Every commercial airplane has at least two completely different sets of sensors, electronics, and software. Both work completely independently from each other and feed the information to the pilots. When both agree, the pilot knows that everything is working properly. But when one fails, we have a trained human deciding which one is failing and which one should be trusted now. Since its operator flies inside the airplane, they have all the incentives aligned for that plane to lift off, fly and touch down as safely as possible. Literally, their own life depends on that. An airplane is a tool being commanded and supervised by the pilot. As long as the pilot has aligned interests with the rest of the passengers and has the last word, no amount of AI will make planes less safe.

The same is valid for any AI. If the operator has incentives properly aligned (the pilot is inside the airplane) and a human is responsible, accountable, and implicated in the final call (supervises), AI will be as safe as any other technology.

Should people use a fork to eat soup?

Of course not. In the same way, no one should use a text generator to solve math problems or generate software.

From Techstory:

"STUDY REVEALS RAPID DECLINE: MATH ACCURACY OF CHATGPT PLUMMETS FROM 98% TO 2% IN JUST MONTHS"

From “How Is ChatGPT’s Behavior Changing over Time?” article:

"Overall, the number of directly executable generations dropped from March to June. […] over 50% generations of GPT-4 were directly executable in March, but only 10% in June. The trend was similar for GPT-3.5. There was also a small increase in verbosity for both models."

The ChatGPT model was created to generate text that looked like a human wrote it, so it learned to do that. By accident, because human text also contains math, it learned a bit about that. But we need to pay attention to what “learn” means in the context of AI. It updated its huge table of probabilities to generate (with high probability) the text “2” when someone prompts it with “1 + 1”. It never understood (as a human) that the context was different and that math has internal rules that must be respected, so if you prompt it for “2 + 2” it may answer 5 or anything else because it is not doing math. It is outputting text with a high probability of being what you want.

The same is true for code generation. Programming languages are a different type of language, and ChatGPT was only trained on that by accident. It never was generating code. It was writing words that had a high probability of satisfying what the operator asked for.

And ChatGPT is definitely (at least until OpenAI trains it with this goal) the completely wrong tool for that. To solve math problems, WolframAlpha is one of the most fantastic math tools and has been around for decades. To write code? Github’s Copilot is trained precisely to do that.

Using a screwdriver to push a nail into the wall may even work, but using a hammer will make things way more effective.

AI scalability

Another good example of AI misunderstanding is how overstated the speed and reach of transformations ChatGPT could bring to the world.

One of the reasons is that OpenAI was a research lab until two years ago. It was not an internet-scale company capable of deploying software used simultaneously by the whole world. Having a model that a small group of people can use is completely different from having a product that the world can use. Even with Microsoft (an internet-scale company) investing 10 billion USD in them, it will still take time to improve the technology to be usable on this scale. Google’s Bard might have a better chance at this if it was not too rushed. Google’s infrastructure is unquestionably battle-tested, and if Bard was built on top of it, it is reasonable to assume that it is substantially more scalable (or easier to scale) than ChatGPT.

In the concept of reach, it's even questionable the idea that ChatGPT was responsible for the most significant technical advancement of AI in the world today. In February 2023, Meta announced LLaMA: the Large Language Model Meta AI, an open-source software that rivals OpenAI's proprietary GPT products. That announcement gave birth to an explosion of AI-powered functionalities. Llama has fewer capabilities than ChatGPT or Bard, but it has the advantage of being open-source. Just some months after version 1 was launched by Meta, the open-source community improved it so much that it was possible to run it on a personal machine without the need for a dedicated server.

Microsoft, an OpenAI backer, also showed signs that they understand ChatGPT's limitations and moved to have a deeper influence in the LLaMa project, which we, in fact, consider the most revolutionary advance in AI at the moment. LLaMA 2 was just announced by Meta in partnership with Microsoft.

And this brings us to the costs of AIs.

AI training and costs

It is no secret that OpenAI spends millions of dollars to train and operate ChatGPT. Bard may be a bit cheaper because Google has been operating on an internet scale for a long time. But we need to dive a bit deeper into what training means in the context of AI.

Since AI is a huge table of probabilities, training is the process of computing these probabilities. Again, the name is misleading. Contrary to human training, there is no repetition to get as close to perfection. What happens (in the specific case of ChatGPT) is that you literally send most of the written content of the internet to this machine that computes probabilities and stores them somewhere. This process is as brute as it sounds and requires humongous processing power.

To illustrate the scale of such costs, most companies that use AI train their models on gigabytes of data. Really sophisticated and expensive models may require terabytes of data. ChatGPT and others require petabytes or exabytes of data. The scale is so huge that it is not replicable by most companies, so they will have to use these capabilities as a service and either pay huge fees to OpenAI or Google or use open-source solutions like LLaMA, which is free to use and able to do most of what Bard and ChatGPT can do.

And Musk again

And if you remember our newsletter on AI and fear-mongering executives, we said:

"In March 2023, Elon Musk returned to the scene by leading a plea for a halt in AI development, leveraging the collective imaginary of a Terminator-like future, and fueling fear-mongering narratives the media and public opinion grew so addicted to."

Now guess who has decided to create a new AI company? From the Financial Times:

"Elon Musk has formally launched xAI, an artificial intelligence company aiming to challenge the dominance of Sam Altman’s OpenAI and “understand the true nature of the universe”."

We could say a lot of things about that, but as Signal’s Meredith Whittaker brilliantly put it in an interview:

"Unlike many other tech entrepreneurs and academics you didn’t sign either of the two recent petitions, the Future of Life institute “pause AI” letter or last month’s Center for AI Safety “existential threat” letter.

No. I don’t think they’re good faith. These are the people who could actually pause [AI research and development] if they wanted to. They could unplug the data centres. They could whistleblow. These are some of the most powerful people when it comes to having the levers to actually change this, so it’s a bit like the president issuing a statement saying somebody needs to issue an executive order. It’s disingenuous."

Our guess is that the AI-pocalypse is not coming anytime soon.

5 current AI discussions: a technical overview