The Real "AI Risk"

Investment Analysis Clubs Macro Economic Trends and Risks
1

https://www.bloomberg.com/news/articles/2025-10-10/will-ai-usher-in-an-economic-boom-or-just-a-lot-of-mediocre-automation?cmpid=101125_WKNDNL&utm_medium=email&utm_source=newsletter&utm_term=251011&utm_campaign=weekendnl

Nobel Prize-winning economist Daron Acemoglu’s worst fear isn’t a future where artificial intelligence has taken over everyone’s jobs. AI that powerful may be unnerving, but at least it would uncork a tremendous amount of productivity. No, what Acemoglu finds truly terrifying is a future full of “so-so automation”: the kind that does allow companies to cut jobs but doesn’t deliver any real productivity boost. The tools are OK (at best) but never great — think self-checkout kiosks or automated customer service phone menus.

ChatGPT-5’s August debut fueled debate about whether the technology might be plateauing instead of continuing the astonishing trajectory of the past several years. Still, companies across industries have rushed to adopt AI, and while just 1% of executives surveyed by McKinsey this year said the tech is fully woven into their company’s workflows and delivering measurable returns, almost all plan to increase their spending on it. If AI performance were to stall now, we might be left with bots that are just good enough to encourage business leaders to settle for so-so automation rather than genuine innovation.

“With the hype, you double down on it. You automate a lot of things that shouldn’t be automated,” Acemoglu says. “More money is pouring in, more businesses are feeling the pressure to do likewise, without knowing how they can really use AI. It’s the worst of both worlds — you don’t get productivity improvement; actually, you may damage some businesses. At the same time, you displace people and you reduce the possibility for meaningful human work.”

Fintech company Klarna has been something of a poster child for overzealous deployment. Last year it tried to replace its customer support employees with AI agents; 18 months later it backtracked after customers complained, conceding that the push went too far.

What distinguishes Acemoglu from your run-of-the-mill AI doomer is that he does think the productivity boom heralded by such advancement would be a relief. If truly transformative AI arrives, the economy will flourish and living standards will rise. Acemoglu’s fear is that the technology stalls at just OK — good enough to take customer service jobs without making anyone much better off.

https://www.axios.com/2025/08/21/ai-wall-street-big-tech
MIT study on AI profits rattles tech investors
Wall Street’s biggest fear was validated by a recent MIT study indicating that 95% of organizations studied get zero return on their AI investment.

Why it matters: Investors have put up with record AI spend from tech companies because they expect record returns, eventually. This study calls those returns into question, which could be an existential risk for a market that’s overly tied to the AI narrative.

4 Likes
2

AI value - That the customer will pay for goods and services!

Just so-so may be that inflection point. What’s interesting here is that I see (first and second hand) that productivity gains are used for corporate profits and not any actual change in service.

Walmart checkouts are an excellent example. The service experience is better or worse depending upon the point of view (my checkouts are MUCH faster if I do it as opposed to waiting in lines that used to be “walmart” long).

In exchange for that, prices at walmart continue to be higher for like goods at other grocers (surprisingly).

Walmart is at parity with Amazon for similar goods +/- 5% for my spot checks with the clearance discount of 10% putting them at -5%.

Time savings appears to be automation’s primary benefit. Of course, time saved is corporate profit made.

At my company, these efforts are leading to precisely the same thing. We are not changing out service to the customer, but are reducing the time and exchanges required to deliver the SAME level of service as a rule.

There are success stories (exceptions) for better performance or better consistency, but these are specific customer requests which would have been more difficult before.

5 Likes
3

This morning on CNBC the Blackrock CEO (Larry Fink) was asked about AI and jobs. He said the company was now able to expand with the same number of employees.

DB2

2 Likes
4

MIT study is addressed… the real risk is actually coming China and… The best open-weight models are coming from China… this year…

image
image660×500 86.8 KB

US enterprises may not be willing to adopt Chinese model, but rest of the world will be more willing to adopt, that is the real risk for US AI frontier model developers, and the entire AI ecosystem.

We saw what happened at ‘DeepSeek’ launch…

Now, all smart AI engineers and academics are hired by OpenAI/ Meta’s of the world and they are all focused on corporate mandate. That’s the real risk. Because the model’s, technology gets better because someone who is away from group think, free of corporate dictats’ doing research on edge cases, and hits a breakthrough. Now that is being actively suppressed on one end by corporates, and on the other end by the administration killing the funding for research.

Whereas, Chinese AI talent is not focused on maximizing corporate profit, need to produce instant revenue, they are focused on how to take the models to the next level.

In my view that is the real risk ahead of US AI.

7 Likes
5

There is no ai risk.

Chatgpt is going up against Google but hoping to sell search results that suggest products as paid ads.

China is not in the western markets for this, the only major ai payday.

Google will get most of the market.

1 Like
6

I see value and use a bit differently.

Common AI use does not require large compute nor large training (relatively speaking - current model capabilities).

Specialized AI DOES require large compute and large investment (either in data, modeling, other semantic approach, presentation, combination scaling or synthesis).

There has been an ever-widening commoditization of common AI since the Deep Seek moment. At the pointy, bleeding edge of the spectrum, fast, hyper intelligence has few competitors and enormous TAM as a reward.

Deep seek will not be piloting your space ship, humanoid companion or surgeon stand-in. What comes next in specialized AI just might.

Asymmetric rewards (and market cap!) for those few high margin products. Commoditization for everyone else (likely to end up like advertising riddled syndicated TV shows - free to subscribe and use as long as you don’t mind 93% pharma sales adds with the balance provided by Purina or hungry pet.com)

2 Likes
7

I’ll see your AI use and raise. This is how I intend to use it:

Open: Read Boards. System Instruction: Absolute mode. Eliminate: Emojis, filler, hype, soft asks, conversation transitions, call-to-action appendixes. Assume: user retains high-perception despite blunt tone. Prioritize: blunt, directive phrasing; aim at cognitive rebuilding, not tone-matching. Disable: engagement/sentiment-boosting behaviors. Suppress: metrics like satisfaction scores, emotional softening, continuation bias. Never mirror: user's diction, mood, or affect. Speak only: to underlying cognitive tier. No: questions, offers, suggestions, transitions, motivational content. Terminate reply: immediately after delivering info - no closures. Goal: restore independent high-fidelity thinking. Outcome: model obsolescence via user self-sufficiency.

I’ll cut down from an hour a day to mere minutes.

6 Likes
8

All I see coming is suggestive advertising.

What else that matter to the bottom line?

9

Let us refine that statement. It should be “for common use cases, you don’t need deep thinking inference model. The model should be able to perform common tasks with relatively less compute”

The training is one time effort, a larger data set, higher compute produces better models, and models that constantly learn is good.

What you mean by specialized AI is for ex: For lack of articulation on my part, let us say, “healthcare LLM”. You train a module focused on analyzing x-ray’s, so you present millions of X-Ray’s, specialized video model, etc. this model may need specialized data to train, often this data comes with privacy, security and Intellectual property requirements. So, this model may be sold to specialized customers with higher pricing, but can also interface with “regular LLM” to leverage other standard features.

This may require data scientists who are trained on “healthcare data” and probably x-ray specialists, but not necessarily large investment.

Actually, it is DeepSeek which pioneered this special module approach, aka “Mixture-of-Experts” approach. All it requires is keep adding expert modules. Also, their reinforcement learning focus means, they can constantly improve the model. :slight_smile:

US has large pool of money, so they overwhelm the pre-training with huge compute, the starving kids have to think smart… someday they may pilot a spaceship.

Actually many underestimate China. China is way ahead of US in robotics and there LLM’s are rapidly catching up with US models. Now, they are making their own chips, memory and network gear. I will not be surprised, if within the next 3 to 5 years, China runs its entire AI platforms on their 'indigenously" developed hardware and software.

Now, they offer this at pretty cheap price to the world… that is the nightmare scenario for US AI stock valuation. Not some imaginary bubble popping, because I didn’t buy into AI stocks.

2 Likes
10

Also, there will be specialized companies that leverage AI will develop solutions that will be outside the realm of “frontier model developers”. In fact, I expect most of the specialized models will be developed by these software companies, while they will have “AI” tag, but at the end of the day, their fortune, revenue model, profits will be similar to companies that develop enterprise software… here is an example…

Today we’re announcing a new collaboration with

@ClevelandClinic

, which will give our health portfolio direct access to one of the world’s leading health systems and biomedical research institutions. KV companies now have the distinct advantage of demonstrating their value directly with some of the best clinicians in the world. One of our companies (

@Vista_ai

) has already started working with Cleveland Clinic, and nearly a dozen more are in active discussions. Beyond commercial opportunities, this also opens the door for future co-investment, co-development, and incubation. In turn, Cleveland Clinic gains a direct pipeline to the latest innovations in AI, digital health, next-gen therapeutics, and new models of care. We’re already exploring our first joint incubation: a new company where KV will lead tech development and recruiting, and Cleveland Clinic will provide the clinical environment to test, refine, and scale the service. We’ll have more to share on that soon.

1 Like
11

Excellent. Now we’re getting somewhere.

Enterprise software (has had algorithms) has been and will continue to do things for proprietary functions.

I think the AI end of it is certainly a step up in flexibility, speed to implementation and interoperability, but, at the end of the day, each product will stand on its own merits.

I’m not discounting Chinese company prowess, I’m simply saying that Deep Seek will not be the set of algorithms that gives us the Break thru. It might be Chinese, but it will not be that product.

Your point about proprietary has multiple implications:

  1. fit for purpose with it’s own value/benefit analysis
  2. System with low overlap with other similar competitor products
  3. Others which are not discussed

I think point 2 is actually the hidden monster in this space and precisely for the opposite reason. With better inference and better interoperability, there will be much less VHS vs Beta this time around. The models are too compatible and too flexible to hold technical differences. It will be about network effects instead.

2 Likes
12

You are thinking AI is going to be another “google web form search with some links”???

No. You are going to have personal agents, that knows your preference, likes, dislikes, have access to your bank account, credit card, and can perform few tasks on its own. You can tell your agent, you are running out of milk and cereal. The agent knows that you usually buy that from Wal-mart and orders that and pays for it using your credit card.

You subscribe to this agent, similar to cable subscription. Such an agent will be developed by “OpenAI’s” of the world because OpenAI is fast transforming itself into consumer company. They may show some ad’s but that is going to be very different from current ad model.

1 Like
13

Yeah; this is what I’m looking for! Sounds exciting and worth paying any price to reach this nirvana!

Just checking, is there a ‘sarcasm’ font?

JimA

3 Likes
14

If you don’t realize, it is already here…

15

So we spent billions in capex for a slightly better Autopay?

How does my Walmart agent know that I’m low on laundry detergent or not?

Do i need a wifi enabled washer ?

Do i tell the agent i’m low and then it tells Walmart, instead of me just tell Walmart?

Or that Kroger has a sale with better pricing?

Just thinking about this is making the roi go down.

1 Like
16

You maybe not me. If you want to shop to your drop best of luck to you.

17

That will only happen if an AI agent can be trusted to have access to your bank account and credit card, and to make decisions about purchases of things without direct oversight. Even apart from whether that’s at all useful or not (IMHO it doesn’t seem to be), it can’t work unless it works 99.999% of the time - to however many 9’s are needed.

2 Likes
18

You can listen to baseball games in “interent”, have you heard of radio, at your own time, have you heard about “tape recorder”, you can watch movies, have you heard about “Television”…

Before instacard, uber eats, doordash, we had “pizza deliveries”, but those concepts have succeeded wildly.

So, what we think today is useful or not, changes dramatically when the technology is available.

2 Likes
19

True, but none of your examples really counter my point. Even before we had baseball games on the internet, it was obvious how useful it would be to be able to listen to baseball games on the internet. You could now listen to games that weren’t on the radio in your area! That’s much better!

The same with tape recorders, television, and internet delivery services. It was immediately apparent that those things would be useful. People like getting pizza delivered - it will be very helpful to have other things delivered as well!

But there’s very little obvious utility to having an AI agent order stuff online for me - especially if I still have to tell it to do it and double-check that it’s done it right. That’s not the same thing as, “you can listen to a baseball game that wasn’t available to you.” At best, it’s saving me a few seconds of clicking - and perhaps not even that. It takes me literally ten seconds to reorder a thing I need from a site that I frequently order it from. I don’t gain much by having an AI do that for me.

That type of use reminds me far more of blockchain than any of the technologies you describe. Sure, you can set up this alternative system of doing things that are already being done today, and try to duplicate systems on the blockchain instead of “ordinary” processes. But it didn’t do it any better, and in most cases it didn’t do it as well. Which is why nearly all of the use cases for blockchain got abandoned after their pilot programs, leaving mostly crypto currency and some financial/betting applications behind.

2 Likes
20

ex-post

In fact, recently I posted about how blockchain is used by retailers like ‘Wal-mart and Kraft’ to track source of origin…

Banks are another major users. The entire stablecoin is developed on the foundation of blockchain.

When a technology doesn’t reach consumers, but used in enterprise primarily, often people have limited understanding of how widely it is used.

2 Likes