Failing to Understand the Exponential, Again - Julian Schrittwieser \ stacker news ~AI

pull down to refresh

Failing to Understand the Exponential, Again - Julian Schrittwieser www.julian.ac/blog/2025/09/27/failing-to-understand-the-exponential-again/

807 sats \ 9 comments \ @Scoresby 23h AI

People notice that while AI can now write programs, design websites, etc, it still often makes mistakes or goes in a wrong direction, and then they somehow jump to the conclusion that AI will never be able to do these tasks at human levels, or will only have a minor impact. When just a few years ago, having AI do these things was complete science fiction! Or they see two consecutive model releases and don’t notice much difference in their conversations, and they conclude that AI is plateauing and scaling is over.

Given consistent trends of exponential performance improvements over many years and across many industries, it would be extremely surprising if these improvements suddenly stopped. Instead, even a relatively conservative extrapolation of these trends suggests that 2026 will be a pivotal year for the widespread integration of AI into the economy:

Models will be able to autonomously work for full days (8 working hours) by mid-2026.

At least one model will match the performance of human experts across many industries before the end of 2026.

By the end of 2027, models will frequently outperform experts on many tasks.

I sympathise with the way he calls out the current breathing spell in AI, where negativity seems to be the most common sentiment. Clearly, the usefulness of this tool is just beginning to seep into our society and we haven't seen anything yet. I have a feeling we will return to the hyperbolic extreme of "AI is going to change everything!!" soon enough. I also appreciated this fellow's gratuitous use of exclamation points, a style I call techno auptimist.

view all related items

256 sats \ 0 replies \ @optimism 16h

I'd like to attack the premise here a bit.

People notice that while AI can now write programs, design websites, etc, [..]

It can only do this in the opinion of people that aren't capable of writing said programs or designing said websites themselves, and we see the newer models not making much headway into being much better.

gpt-5 regressed against gpt-4o in instruction hierarchy - meaning that important instructions are ignored more often - and this has, despite being promised to be mitigated, in almost 2 months not been mitigated according to their own, current, system card. Especially cognitive rulesets, which is what we use in coding instructions, but also in the detection of malicious user messages on chatbots, regressed according to OpenAI themselves from a 73% success factor to 62% in gpt-5. I've lightly observed similar outcomes in claude-4-sonnet vs claude-3.7-sonnet in a vibe coding setup.

So it's not a "breathing spell", imho. It's a failure to deliver an improved product. Products that had room to grow, like qwen-2.5 to qwen-3, have in the meantime improved, although that mostly means that they have caught up with the bleeding edge a bit.

There are however 2 products that did deliver some improvement: Gemini 2.5 and Grok 4, but they too have slowed down a bit.

Remember that errors compound in agentic setups because those are loops. So if you have a 5% error rate on one-shot, then you have a ~23% error rate when you have the same LLM perform 5 steps on top of each other.

152 sats \ 3 replies \ @SimpleStacker 23h

Yeah, it is funny that people are saying that AI hit a wall when it's only been 4 years since we were even impressed that it could have a conversation.

I'm definitely an AI bull (in the business sense), but I am pretty concerned about what it could mean for society. Certain technologies may not have been a net positive (smartphone comes to mind).

110 sats \ 2 replies \ @Scoresby OP 23h

How would we even go about measuring the net effect of smart phones?

A lot of us (myself included) probably feel like we spend WAY too much time looking at our phones. But then again, prior to the printing press, pretty much nobody looked at books. That change took longer, but comparing the amount of time a person in 1990 spent reading to somebody in 1300, it might feel like a pretty big shift, even an unhealthy shift. I'm sure that there's a case to be made that books have changed humans in inhuman ways.

So, yes, smart phones are unhealthy, but they also come with a lot of benefits (instant access to all the information on earth + ability to talk to anyone anywhere anytime + google maps + ...).

AI is probably gonna feel like smartphones in that it will seep into our lives way more than we are comfortable with.

Also, there's this: I don't know how to resist strong technological change. I spent a lot of the first few decades of my life resisting technology. But looking back now (from the vantage of the smart phone I'm clutching in my sweaty hands), it seems like my resistance was futile at best, wrong-headed at worst.

178 sats \ 1 reply \ @SimpleStacker 22h

Yeah I don't think there's a real way to measure that, since everyone disagrees on what makes for a good society anyway.

I just think that for me, personally, I'm willing to give up the benefits of the smartphone to return to a society where everyone was more present, focused, and engaged, versus distracted by their phones all the time. Especially when it comes to the youth.

We can still communicate on the go from anywhere (dumb phone era). We also had access to interactive maps (GPS systems). Those to me were the two most useful features of smartphones that i'd want to keep, but they were already available.

166 sats \ 0 replies \ @Undisciplined 21h

I don't think it's really an issue of the technology being a net negative as much as it's an issue of social norms not having adapted to the new technology yet.

102 sats \ 0 replies \ @south_korea_ln 12h

In the current paradigm of LLMs, I'd put my money on it reaching a plateau with small incremental changes.

You can't just extrapolate the last few years and assume it'll keep going.

The next jump will imo only happen with a whole new fundamentally different type of approach.

0 sats \ 0 replies \ @Solomonsatoshi 17h

Moores Law is in gradual decline.

0 sats \ 0 replies \ @brave 19h

This nails the disconnect between AI’s current flaws and its insane potential. The exponential trend is hard to argue with, look at how fast we went from basic chatbots to AI that can code and design. But I think the skepticism comes from people wanting AI to feel human in a way it’s not yet.

0 sats \ 0 replies \ @035736735e 22h outlawed

stackers have outlawed this. turn on wild west mode in your /settings to see outlawed content.