Interview: A new paradigm in AI

February 10, 2025

Good morning. This Super Bowl Monday marks the beginning of France’s AI Action Summit, an international gathering that will seek to explore some forms of global AI governance, among other things.

And, as always, there is so much going on in this space that it is quite literally impossible to keep up. So today, we’re taking a deeper look at just two stories — the first, the state of the AI trade. But the second … a whole new paradigm.

— Ian Krietzberg, Editor-in-Chief, The Deep View

In today’s newsletter:

💰 Is the AI trade still intact? Here’s what we know so far
👁️‍🗨️ Interview: The start of a new AI paradigm

Struggling with work-life balance in 2025? It’s time to change that!

Join this 3 hour Intensive Workshop on AI & ChatGPT tools (usually $399) but FREE for first 100 ‘The Deep View’ readers.

Save your free spot here (seats are filling fast!) ⏰

An AI-powered professional will earn 10x more. 💰
An AI-powered founder will build & scale his company 10x faster 🚀
An AI-first company will grow 50x more! 📊

🗓️ Tomorrow | ⏱️ 10 AM EST

What you’ll learn:

✅ Make smarter decisions instantly with AI

✅ Automate tasks & boost productivity

✅ Scale your business faster using AI

✅ Save $$$ by simplifying complex problems with ChatGPT

👉 Hurry! Click here to register (FREE for First 100 people only)🎁

Source: Unsplash

Wall Street has not been overly pleased with the bulk of the megacap tech firms, lately.

What happened: Since reporting earnings, Google is down 10%, Amazon is down 3% and Microsoft is down 7%. Meta’s stock — up 2% — is the only one among them that hasn’t fallen since reporting earnings.

It’s not clear that there’s a specific culprit behind the Street’s negative feelings toward Big Tech at the moment (Tesla is also down 10% since reporting earnings), although each of these companies reported earnings after DeepSeek’s release of R1, a moment of reckoning for the industry.

The other theme that is permeating across these hyperscalers is a simple one of, quite frankly, insanely high expenditure. Collectively, Meta, Microsoft, Amazon and Google told investors that they plan to spend around $320 billion in AI-related capital expenditures in 2025, a significant increase over their collective $230 billion in total capex from the year before. (That’s $100 billion from Amazon, $80 billion from Microsoft, $75 billion from Google and $65 billion from Meta).
The messaging from each of the companies is (and has been) the same, perhaps best summed up by Amazon CEO Andy Jassy: “I think that both our business, our customers and shareholders will be happy, medium to long-term, that we’re pursuing the capital opportunity and the business opportunity in AI,” he said last week. The view from the hyperscalers is the investment will pay off in a big way, just give it some more time.

This, rather conspicuously, despite the impressions roiling off of DeepSeek, that AI developers can conduct final training runs of state-of-the-art models without excessive quantities of the compute the hyperscalers have been so busy acquiring.

This is also despite the fact that return on investment remains minimal compared to the scope of the expenditure, a point that Wall Street has been nervously observing for about a year, now.

But to DeepWater Management’s Gene Munster, the capex boost “underscores that we are still in the early stages of both the AI hardware buildout and the AI software trade.”

Shares of Nvidia are up around 2% since Big Tech earnings season began.

Munster said that, if DeepSeek hadn’t emerged, the market would be celebrating the capex boost; the fact that it’s not is proof that investors are “uneasy” right now, “given the massive rally over the past two years and the lingering fear that those gains could quickly unwind if hardware growth slows.”

That fear has become so deep-seated (pun intended) that Munster doesn’t even think that Nvidia’s late-February earning report will convince investors that “the coast is clear.”

Despite that, the capex is clearly in place, which to Munster, means the AI hardware trade is still intact, for now.

If you study the capital flows and announcements of the world’s largest asset managers, you’ll notice that private credit has been one of the hottest asset classes of 2024. Historically, private credit has outperformed both stocks and bonds in the last 3 market downturns.

With Percent, accredited investors like you can access institutional grade deals with:

Higher Yield Potential: Percent's average net returns to investors in 2024 was 14.9%.
Short-Term Flexibility: Deals on Percent typically mature between 6 and 36 months, with an average term of just 9 months.
Diversification across sectors like small business loans, consumer finance, and real estate throughout the U.S., Canada, and Latin America.
Low Minimums: You can get started with as little as $500

Investors turn to private credit because of its potential for returns and stability. In the last 3 market corrections, private credit outperformed both stocks and bonds.

New to private credit? Earn up to $500 when you make your first investment.

OpenAI’s newfound partner Softbank is reportedly close to finalizing a $40 billion investment in OpenAI at a $300 billion valuation. If it goes through, Softbank would leapfrog Microsoft to become OpenAI’s biggest backer.
The UAE has teamed up with France to build a massive, new AI campus in France, a partnership that could represent “30 to 50 billion euros” worth of investment.

Elon Musk’s DOGE is working on a custom chatbot called gsAI (Wired).
Sutskever’s SSI in funding talks at a $20 billion valuation (Reuters).
Meta launches new program to improve speech and translation AI (TechCrunch).
Anduril in talks to raise money at $28 billion valuation as defense-tech booms (CNBC).
Federal judge temporarily blocks DOGE from accessing sensitive Treasury Department payment systems (NBC News).

Source: Created with AI by The Deep View

While the latest advances in generative AI tech are proving that benchmark tests aren’t that hard, after all, they have also reaffirmed a few fundamental flaws of the architecture.

Let’s take OpenAI’s Deep Research, for example, an “agentic” system that combines OpenAI’s latest o3 “reasoning” Large Language Model (LLM) with a mechanism for real-time web search. As much as some early users, such as Wharton professor Ethan Mollick, remain impressed with the system, saying it's unlike other AI releases in terms of real-world impact, others — Mollick included — have noticed something: it still gets information factually wrong.

Only now, these hallucinations are buried in the midst of impressive-looking, lengthy reports, making them a little harder to catch unless you look closely.
For Mollick, it doesn’t matter if the system is right or wrong; the utility of Deep Research is as a means of prompting and enhancing his own thinking. And for that, he says, it’s more than good enough.

Whether people find value in factual inaccuracies or not doesn’t really matter. The point is that, more than two years since the first version of ChatGPT sparked a global AI race, we’re still talking about the same fundamental limitations. Hallucinations haven’t gone anywhere.

This is something cognitive scientist Dr. Gary Marcus has been talking about for a long time; hallucinations, he has said, are a fundamental flaw of LLM architecture. The only way to beat them is to dump LLMs entirely or incorporate them into a new system, a better system. A hybrid system.

A neurosymbolic system.

I talked to the team at Imandra that’s building one.

But first, what is it? As we talk about often here, the unfortunate term “AI” functions as a vast umbrella term, encompassing many different technologies. Within that umbrella is something called symbolic AI, sometimes referred to as GOFAI (good old-fashioned AI).

Here’s how it works: symbolic AI relies on explicit symbols to encode and reason about knowledge. This is a rules-based, deterministic approach to mimicking intelligence (think ‘if, then’ statements). It is notably explainable and transparent, meaning it’s easy to see how a symbolic system arrived at its conclusion.
But the recent AI craze represents a push away from symbolic AI and into deep learning, which works like this: powered by artificial neural networks — designed to mirror the human brain — deep learning is trained to analyze and uncover patterns within vast troves of data, without being explicitly programmed to do so.

But deep learning is a bit of a black box; no one knows how LLMs arrive at their output, but what we do know is that, while highly scalable, they’re brittle — a lack of rules makes them unpredictable, and they’re further vulnerable to biased or bad data and prone to outputting false information (hallucination).

The promise, according to Imandra co-founder and co-CEO Dr. Grant Passmore, is to combine the best of each of these worlds into a hybrid system whose “reasoning” is grounded in the mathematical, rules-based logic of symbolic AI, but whose communication is powered by the adaptability and statistical power of language models.

Passmore teamed up with co-founder and co-CEO Denis Ignatovich in 2014 to launch Imandra, an AI firm built around the Imandra automated reasoning platform, a system Passmore created to ensure trust and reliability in complex algorithms. The system is leveraged by a number of institutions and organizations, including Goldman Sachs and Citi Bank. Imandra has raised a total of around $12 million in funding, according to Crunchbase, most recently in 2021.

The plus side of LLMs, Ignatovich told me, is that they’re really good at translation. The downside is that LLM ‘reasoning’ “lacks rigor.” Symbolic AI, on the other hand, is “incredibly powerful,” but requires a precise level of expertise that makes them more difficult to use.

“Once you combine them, now you have this very easy-to-use interface that can translate from English prose into precise mathematics … which has been a complete game changer,” Ignatovich said.

Imagine two gears rotating against one another, working in tandem. The language model ‘gear’ can, according to Imandra, “delegate complex reasoning tasks” to the symbolic ‘gear,’ enabling output that’s rooted in “the laws of logic.”
All the while, the LLM component of the system is able to flexibly translate the precise, mathematical terms of the symbolic component, making for a fluid and accessible system.

A part of this interconnected system enables visible, auditable traces of the symbolic reasoning that occurs, which “allows you to not only solve harder problems, but to get out results with evidence that can then be independently examined using logic, as opposed to just a massive text where real laws of logic may not have been followed, and then you have no way to automatically vet its veracity,” Passmore said.

Due to its grounding in the logic-based approach of symbolic AI, most classes of hallucination errors are eliminated automatically, according to Ignatovich. And due to the explainability baked into its translation of formal mathematical languages, any hallucinations that do occur are easy to spot as logical failures.

Though this field of automated reason is a bit obscure, especially compared to the current, raging popularity of deep learning, it’s an approach that is gaining some steam. Amazon, according to the Wall Street Journal, is making a big bet on automated reason, specifically as a hedge against the hallucinatory limitations of LLMs alone.

“For those on the computer science side, it's known,” Passmore said. “LLMs cannot learn arbitrary reasoning. And so what do you do? Well, you have to bring in the symbolic reasoners.”

He said that this neurosymbolic approach is “100%” the trajectory that the industry is on.

“It's just a matter of time (until) everyone recognizes, okay, you have to integrate the symbolic side. And so that's what we're trying to make all the infrastructure for,” Passmore said.

It’s an approach that the two think will be a critical enabler for the wave of ‘agentic’ systems that are currently on the rise.

The interesting thing about symbolic AI is that, according to Ignatovich, it runs on CPUs, the older, far less energy-intensive cousins of the GPUs that LLMs are built on.