Take a Deep Breath

A week is a long time in AI it seems.

Right after I posted last time, DeepSeek-R1 (called “DeepThink” in their app) was released, the $500bn Stargate initiative was announced by the US President (I wonder if OpenAI have the IP rights for that!?), and OpenAI launched Operator. So I guess I’m lagging behind and you may well have been flooded with insights already. That said, I’ve had a couple of people ask my thoughts so that will be today’s focus.

But first, since I wrote it last week and committed to sharing how I am thinking about the start to 2025. I hope they’ve aged reasonably well!

???2025 Opinions

OK, some relatively quick fire (and hopefully a little contentious) views on where were headed and what we’ll see in 2025.

Upskilling

As I alluded to in the intro, this is the year of upskilling. All “white collar” jobs will change, if they haven’t already. Some will be subtle changes, but everyone needs to learn new skills. At the risk of preaching to the converted, if you’re not already learning about and using AI tools, I’d highly recommend considering this a New Year’s resolution to yourself.

In Jun 2023 I wrote a short somewhat tongue-in-cheek blog called “Replacing APIs with Middle Management” which was an inversion of a somewhat viral piece written in 2015. The point I was trying to land was with natural language poised to become the interface for all computing systems, exceptional prompt engineering seems a lot like one of the jobs of a good manager.

So here’s the provocative opinion - we’re all going to be “managing AIs” at some point in the near future, even if that AI is your personal assistant, because removing some of the daily drudgery will significantly improve productivity. These are skills we’ll all need to learn for the future workplace.

AI Adoption

The corollary here is that this is the year of AI adoption in the enterprise. For the most part, 2024 was about pilots and experimentation and it was also the year of people bringing their own AI. In 2025, AI is high on the agenda of most CEOs, which means it is going to be a feature of most enterprise leadership teams OKRs, which will naturally include HR and L&D teams. Josh Bersin is describing this as “The Rise of the Superworker”, and is in part has influenced point 1 above.

Agents

I’m calling it, this is NOT the year of Agents. At least, not in the way I think they are being hyped. Systems that logically reason, plan autonomously, and use tools, to accomplish goals (all without a human-in-the-loop) will only appear in extreme edge cases if at all. Human systems

Fun fact: one of the preeminent books on AI, “Artificial Intelligence: A Modern Approach” by Stuart Russell and Peter Norvig (A Pearson title I might add!) used the agent metaphor back in 1995 (!) when it was first published. Chip Huyen recently wrote a very comprehensive article on Agents based on existing literature, in which he says “AI-powered agents are an emerging field with no established theoretical frameworks for defining, developing, and evaluating them”. So it’s certainly open to interpretation.

For me, what is being “rediscovered” as Agents now is effectively Intelligent Automation (IA), which itself was the evolution of robotic process automation (RPA). And this will be quite significant (I refer you back to opinion #1).

Don’t get me wrong, the significant technological advancements of the last few years are real and impressive, agents in the complete sense will happen, and innovation teams should be experimenting with “agentic” prototypes, but for the core business it wont be this year.

AI Services

An area I am particularly bullish on for growth in 2025 is what I think of as AI services. This is essentially taking expensive processes that are important enough to exist, but are suboptimal because they are impossible to scale. AI in the form of the advanced Intelligent Automation I mentioned is becoming capable enough to solve this, particularly with the potential of reasoning models (more on that shortly).

When the UK PM talked about the AI Opportunities Action Plan, improving public services was a key component of it, I think of this as AI services - behind the scenes quality of life improvements that reduce costs and improve efficiencies. I am convinced this is an area of significant value across education - improving learning outcomes might be about optimising the system so teachers can spend more time with students.

Robots

Those AI employees might well be embodied… I think we’ll continue to see growth in robotics. Since we’ve exhausted training data, we need to find new sources… and robots offer the potential to do so with new sensors and experiences to help build ”world models”. To be honest, I think this could well be on the path to real agents… as with EVs, gathering training data in the real world has been an important part of the road to (no pun intended) autonomous vehicles.

?? DeepSeek

Since there is a lot in the press, I’m going to try to offer my consolidated thoughts. First a couple of recommended readings:

1. Ben Thompson’s Stratechery “DeepSeek FAQ” post.
2. The Forbes article “All About DeepSeek” is a nice breakdown of some of the tech nitty gritty.

So a quick word on the stock market volatility. It seems DeepSeek rushing to the top of the App store may have something to do with it:

The implication here being AI models have no moat, since DeepSeek was able to train R1 on “less that $6M”, release it for free and open source, and that it outperforms o1 pro (the $200 monthly subscription one) on most benchmarks. But if you read the Stratechery post, that the $6M number relates specifically to the final training run only and excludes all other costs. It also appears to relates to Deepseek-V3 (as opposed to R1). Additionally, it’s a “distilled model”, which essentially means it used a process of extracting understanding from another model during training. Those of us who have been following developments have known for some time that intelligence is being commoditised - costs have plummeted by a factor of 1000 the last 3 years. So I don’t think we really ever believed Models had a moat.

Nevertheless, what they've been able to do is incredibly impressive. True innovation really, given they've had to figure out how to do it with obstacles (i.e. less performant GPUs, due to trade restrictions) that perhaps led to architectural breakthroughs. A lesson in unintended consequences perhaps.

I’ve played with the DeepSeek browser app and I have to say it seems very impressive. There are a quite a few posts about the restrictions and censorship that I’m not surprised by. But I don’t really want to focus on that, because the free app isn’t really the story, it’s that the model weights have been open sourced and their research shared. I can’t say for sure, as I haven’t run the model locally, but I suspect the security/safety filters are only applied at the consumer application level not baked into the model.

NVIDIA were a big casualty of the stock market uncertainty but, for what it’s worth, I think everything is up in the air right now and no one really knows what to do. Also, hardware is hard, so I'm holding judgment on NVIDIA for the time being! So I think calmer heads will likely prevail unless you’re at the bleeding edge, then it’s probably a bit uncomfortable. We’ll touch on the geopolitical stuff with Stargate, but I thought I would add this isn’t the end (yet). Sam will likely go full “founder mode” towards his AGI ambitions, probably learning from what DeepSeek have achieved.

But does all of this really matter? Maybe if we believed that intelligence was a zero sum game, but I’d argue that intelligence is subject to Jevons paradox and as supply increases, it expands the frontier of what’s possible, and as such demand increases.

What I do think is that reasoning models will become increasingly important in education. It can engage students in conversational, inquiry-based learning, adapting questions and challenges to their level of understanding, and evaluate open-ended responses (e.g., essays, problem-solving processes) with near-human-level accuracy and provide constructive, contextual feedback.

Unpopular opinion. We need to think about learning from first principles for the AI era. It's hard, because in some respects this could be the innovator's dilemma at an industry scale. Let me ask you this: what does education look like when everyone on the planet with a mobile phone has a free tutor available 24x7?

??Stargate

Alright, let’s keep these last 2 items brief.

Stargate as I mentioned in the intro is a $500bn project focus on AI infrastructure that promises to create “hundreds of thousands” of jobs in the US and “secure American leadership in AI.”

The thing that I found most interesting is that this is an infrastructure project for OpenAI. The new company is a joint venture between SoftBank, Oracle among others, with Microsoft involved as a tech partner. There were some interesting social media interactions and media reporting on whether or not the money was actually available, but Satya Nadella confirmed he was good for his $80bn so I think that probably settles the matter.

I think it also goes without saying that the DeepSeek release has rattled the US, which makes Stargate far more likely to move forward at pace.

A final thought. This raises questions on energy production for such an ambitious project and with lots of talk of nuclear over the last few years, I expect this will increase investment into Small Modular Reactors.

???Operator

OpenAI made their first official move into “Agents” (cough, cough), but for now the beta launch of “Operator” and only available to US-based ChatGPT pro (the $200 subscription).

Check out the official OpenAI video to learn more:

As always, thanks for reading.

Jody

News I Didn't Use

• How has DeepSeek improved the Transformer architecture? (Epoch AI)
• The State of Conversational Voice AI in Education (ElevenLabs)
• The power of personalized AI (OpenAI)
• Buzzy French AI startup Mistral isn't for sale and plans to IPO, its CEO says (Yahoo Finance)
• Which AI to Use Now: An Updated Opinionated Guide (One Useful Thing)