text originally published in Portuguese
What to expect next year?
It is very difficult to make predictions, especially about the future.Niels Bohr
I usually direct my study based on what I believe will happen in the coming years, in a very flawed attempt to predict what might happen to the technology market, with a nice educated guess.
From what I saw this year, I believe we can get an idea of how the market will continue in 2025, and what we can expect from the year.
As in 2023 with GPT, and 2024 with "AI for everything" and the growth of multimodals, the next year will be full of hype and dreams of AGI, but with many more examples of the good and bad of AI models.
First prediction: We won't have AGI in 2025
How I see 2025 from a technology and AI perspective:
I intend to write about each of the topics throughout the year, but it is important to have a list that will serve as a parameter for my understanding and vision of the market during the year.
The intention is not to exhaust the possibilities, just to point out a path, completely biased by my own choices.
Hype Machine has a new target - Test-Time Compute
Models like o3 , Claude 3 , Perplexity AI will be the focus of the entire first half, possibly the year. It is a major shift from the Train-Time Compute paradigm (spends $ to train), to Inference-time Compute (spends $ to get the answer) or Test-time Compute . The subject is interesting and I have been delving into it. It is worth understanding better. This will impact all existing models.
The results presented by OpenAI on o3 are indeed impressive. Test-time compute is a real performance leap. Despite this, a model was presented that spent 1000usd and 15min per attempt on the prompt. 300k usd spent on the announcement only on compute! Unfeasible and probably of no real use for clients that do not have infinite money (like the American army). Even with suspicions of benchmark contamination and no external evaluator, you can expect a marketing locomotive in the coming months.
GenAI Taking Over Content Generation
It’s already happening, and it’s only going to get worse: generative content on Instagram, Soundcloud, Spotify, YouTube and TikTok will impact the entire business model of social media engagement. We’ll see the first moves to contain or differentiate what’s produced by humans.
Search for any topic online about AI, crypto, or tech products to test it out: you’ll invariably end up on channels with AI-generated videos that regurgitate platitudes and mediocrity without any real value. Here’s an example:
Tools like HeyGen, CharacterAI, NotebookLM, Whisper, and many others will become even better, without being uncanny. Rapid consumption of content will be overtaken by channels farming engagement.
The Dead Internet Theory is increasingly becoming a reality. If you work and depend on content generation, I'm sorry.
"Micro Models" with specialized use. Small Language Models (SLMs)
It is a fact that large-scale LLMs have a limit, and that we are getting close to it, hence a concrete shift towards finding other approaches to expand the capacity of these models (RAG, IFC, etc.). As long as you are still using products with the cost subsidized by investors and Cloud Capital owners, that is fine. But that may change next year.
Performance and Cost determine the large-scale adoption of LLMs, and in this regard, the viability of the cost is very questionable. Small Language Models (SLMs) are a very viable possibility.
With the evolution of performance of models with a few billion parameters, we will see more solutions using specialized models for your demand: legal analysis, computer vision, journalistic writing, video generation.
Most Expensive LLMs
Some predictions:
200 usd/month is a test for OpenAI to move forward in the quest for profitability. We will see more price changes across all platforms
o3 will introduce even more MRR upgrade options for the enterprise, which is why the latest demo demonstrates the performance of models 1000x more expensive than the current price
Competitors will follow suit. Expect a significant search for differentiation of pro , premium , ultra models for all vendors.
Even with economies of scale, reasoning models are unfeasibly more expensive for the public, and will be used in part by large companies to leverage infrastructure sales - "with more compute you can solve any problem! 🤡"
Open-source FTW
Since the market is quite dynamic, it has been a challenge to keep up with everything that has been happening. But open source models, and especially Open Weights, have become a huge possibility for using solutions.
And in this regard, the Chinese DeepSeek is one of the most promising, competing head-on with the best current models. In addition, it is really demonstrating results to reduce the training cost, and publishing everything on github . Open source may be the way to win in this market.
AI Agents whether you like it or not
2024 was the year of speculation about agents. 2025 will be the year of medium-scale implementations. Maturity in Governance and orchestration tools can drive experimental use for real solutions.
Specialized media has been talking about the story for some time now, and I hope to see more cases, especially more complex and substantial ones. Now it's the turn of the big channels to talk about it.
Risks inherent to the model and potential problems will be treated with greater scrutiny (we will see more cases of people losing their lives or being deceived)
Giving decision-making power to agents is a mistake , especially with the security risks I mention below.
Security and impacts
Study security for LLMs and Agents. The integration of tools with internal mission-critical systems or confidential data pushed by big techs hides an inherent risk in the LLM architecture. We will see this in practice in 2025.
Add in decision-making capabilities assigned to agents (like Claude’s desktop automation ), and you have a recipe for disaster.
What particularly fascinates me are AI Worms , and the ability to infect an LLM just by interacting with the prompt, and be able to impact any other user in the sequence. Zero-click Worms that leak information or lie to users are already a reality.
Just like the case of Air Canada and Sewell Garcia , we will see a multiplication of cases throughout the year.
Regulation, Regulation and Regulation
I imagine no one wants to live in a world where their digital interactions are through machines pretending to be humans, and the only way to ensure that is through regulation. Regardless of your opinion on regulation, we have a whole year to prepare for the work being done in the world.
The fear inherent in open-source models, due to the possibilities that a model capable of generating convincing images and texts, will push negotiations towards the use of AI solutions
By 2025, as OpenAI grows, we will see more lawsuits against unauthorized use of intellectual property than any other business model. I hope we can compensate artists who have lost a lot of money in this process.
Brazil's AI Regulatory Framework . Although it is weaker than the initial text, the Framework is now awaiting approval in the Chamber of Deputies, and we will probably see the result in 2025. There is still a lack of a clear position regarding violations of the framework, and the eventual penalties that companies will be subject to. But it is a start.
In addition to Europe, I believe that other countries and economic blocs will move in this direction. But real results will only come if they are driven by the financial loss that Studios, Media Vehicles and Large Companies believe they have suffered. Unfortunately.
Medicine and Health
The Baumol Effect puts healthcare in the crosshairs of any technology company looking to leverage profits. Also known as Baumol's Cost Disease, despite the small gain in productivity over the years, salaries have always risen: nursing and medicine, as a profession, continue to be very close to what we have been doing for the last 50 years.
More examples like UnitedHealth, suicides of people using LLMs, misdiagnosis and indiscriminate use of transcription tools will cause serious problems.
My hope is that we will see the first attempt to hold companies accountable for incorporating AI tools into healthcare management in healthcare. Until technologists are held accountable for AI mistakes, we will have a hard time making a positive impact on the market.
Digital Crimes
The greatest achievement of Artificial Intelligence is the ability to deceive and steal from people. No other technology is as productive in the service of criminals. 2025 will be the year of AI-powered digital crime .
Familiarize yourself with the Zero Trust approach to any digital interaction .
Deep fakes are becoming the norm. Without greater regulation and accountability, we will see mainstream media using the technology in advertising, normalizing its use even further.
Digital scams are already common, and they will become even more indistinguishable from reality for ordinary people. Your uncle from the countryside will receive a call from you, with your voice, asking for money, on a scale unimaginable.
Text, image and audio at the service of fake news. It already happens today, and the scale of use is only accelerating
Cases in schools and the use of Deep Fakes to create pornography using images of teachers, students and staff will be more common.
Subsidized tools, open source models and technological advancement will make this topic one of the main ones in 2025
Where we can see real results
More companies will follow Bluesky’s lead in declaring that they do not use web content to train AI. In addition, more products will take a stand against using generative AI. The rising tide of user dissatisfaction will force companies to take a stand.
I believe we will see the main impact, and it will probably be the first real wave of layoffs: BPM with AI. Process automation, first-level service and especially customer service will be impacted:
Impact on Jobs, especially in the creative sector - which is already happening with automatic image generation.
RPA for accelerated Data Entry: solutions that solve general information registration will begin to be implemented
Copilots and consolidation of general tools for development
Use of Computer Vision for security at more local levels (an interesting area to invest in)
Despite my aversion to using GenAI for content (my texts tend to be more in-depth than an AI can deliver), it is undeniable that news, institutional, blogs and general content indiscriminately use these tools for action. The result is there...
If you've made it this far, I hope you enjoyed your reading. We'll see you again in December 2025 to see if anything here made sense and happened.
Opmerkingen