AI’s Dial-Up Era. The essay argues that AI is in its dial-up era for two reasons:

1. AI is still in its infancy stage, where the benefits brought by the technology have increased productivity and jobs, rather than decrease/eliminate it. Looking at other industries, it’s because there is still demand and its growth hasn’t been saturated.
2. AI is showing parallels to the dot-com crash, where a lot of companies jumped on the WWW hype. As the author points out, a lot of companies are becoming overvalued and do not show any Product Market Fit.

The article shows that Automation hasn’t reduced jobs (apart from the tech sector, but that’s a different story) for several reasons:

• real-world complexity
• regulatory/insurance hurdles
• Jevons Paradox: “economic principle that a technological improvement in resource efficiency leads to an increase in the total consumption of that resource, rather than a decrease”.

ByteCoder:

I feel like we’re back in the mainframe era. A lot of software can’t operate without an internet connection. Even if in practice they execute some of the code on your device, a lot of the data and the heavyweight processing is already happening on the server. Even basic services designed from the ground up to be distributed and local first - like email (“downloading”) - are used in this fashion - like gmail. Maps apps added offline support years after they launched and still cripple the search. Even git has GitHub sitting in the middle and most people don’t or can’t use git any other way. SaaS, Electron, …etc. have brought us back to the mainframe era.

bccdee:

I find the argument for the bubble to be extremely straightforward.

Currently, investment into AI exceeds the dot-com bubble by a factor of 17. Even in the dot-com era, the early internet was already changing media and commerce in fundamental ways. November is the three-year anniversary of ChatGPT. How much economic value are they actually creating? How many people are purchasing AI-generated goods? How much are people paying for AI-provided services? The value created here would have to exceed what the internet was generating in 2000 by a factor of 17 (which seems excessive to me) to even reach parity with the dot-com bubble.

airspresso:

I keep hearing that LLMs are trained on “Internet crap” but is it true? Karpathy repeated this in a recent interview, that if you’d look at random samples in the pretraining set you’d mostly see a lot of garbage text. And that it’s very surprising it works at all.

The labs have focused a lot more on finetuning (posttraining) and RL lately, and from my understanding that’s where all the desirable properties of an LLM are trained into it. Pretraining just teaches the LLM the semantic relations it needs as the foundation for finetuning to work.

ChatGPT Atlas is an Anti-Web: The Browser That’s Anti-Web. I’m always late to the party, so I haven’t tried out Atlas. I will also never will because I’m pretty anti-Chrome.

But Anil Dash gives it a good review:

• “searching” will take you a chat session, not the search engine results. Reading this made me realise how much we have taken the URL box for granted. It’s become default UX behaviour to expect search results to appear. Anil searched with the keywords “taylor swift showgirl”. I appreciated that search engine have become so good at recognising intent, because there are various branches to explore from there. Are you looking for informationa about the album? Do you want to buy tickets or merch? Or do you want to read a review? An AI search engine will have to reimagine all these things, which I never had to think about.
• The UX is a mystery and the only way to discover it is by typing. And typing. And typing. Anil gives the example of Zork. You pick up a rock. You move left. In the OpenAI demo, the team member typed search web history for a doc about atlas core design. It goes back to Point 1, where the AI cannot recognise intent so it has to be clearly spelt out. Also, Anil points out that there is a high risk involved if you forget the correct set of magical incantations, because LLMs have a huge risk of hallunciating.
• ChatGPT is very obviously hoovering up data. You are the customer. Warning labels are required because regular humans cannot be trusted to use the technology without self-harm. Agree.

Your Doctor’s Screen Time Is Hobbling Health Care

localghost3000:

I worked in health care tech for about 5 years. AI driven before it was cool. Took processes that normally took years down to a couple hours. Cutting edge stuff.

What struck me over the years was the open hostility we faced from the staff. The admins would buy our product, then have us come do trainings. The clinicians seemed to resent every second of it and would just never use the tool.

Towards the end of my tenure there, a PM said to me “the last thing these people want is to have to learn yet another workflow”. Which is when the penny dropped for me that our tool was just one of a bazillion being force fed to these poor people. They want to spend their time with patients not a screen.

Despite it being the most mission driven I have ever felt about a product (we were literally trying to help cure cancer lol). I’ll never work in health care again. Like education, it’s a quagmire.

Dr James Yacyshyn:

Ultimately, everybody in this conversation agreed that there was value in learning and mastering foundational knowledge. Foundational knowledge was not viewed as just a preliminary step, but viewed as a cornerstone of professional growth and development. However, how we defended this point varied.