In our last edition, we summarized some of the major developments and trends from 2024. For this edition, we wanted to cover our own non-hyped predictions for AI from a business, policy, and technical angle.
In today’s edition:
We will go from pre-trained to tailored solutions
It will be an exciting year for multi-modal models
There will be no red or blue state approach to AI
Leveraging Federal Agency authorities to regulate AI will be difficult under the Trump Administration
Agentic AI will not be adopted at scale within enterprises… yet
1. We will go from pre-training to tailored AI solutions
The era of traditional pre-training (aka training larger models on more data) is ending. We’ve seen rumblings of this in late 2024: Ilya Sutskever discussed this in his late 2024 NeurIPs keynote talk; and the internet is becoming increasingly filled with synthetic (LLM-generated) data that can lead to model collapse when used indiscriminately. While high-quality synthetic data (e.g. in Phi models), reinforcement learning (e.g. in O1 models) and inference-time reasoning techniques will evolve, we believe that industry/domain-specific models will come to the forefront, especially for high-risk and enterprise domains. We have seen this pattern previously: before the breakthroughs with GPT models, domain-specific versions of the, then state-of-the-art, BERT models were popular.
While fine-tuning capabilities already exist, they require resources and expertise, as well as introduce increased risks (like compromised safety and data poisoning). An enterprise-grade “medical LLM” could help a hospital create an AI system that assists with billing (via knowledge of SNOMED and ICD-10 codes) with a much lower lift. From another angle, most LLMs have worse performance outside of a handful of popular languages, global companies may need to build chatbots that specialize in other languages and dialects. The providers of these solutions will be responsible for curating tailored datasets and red-teaming the final systems for risks introduced via fine-tuning. Many enterprises have had challenges transforming prototypes into production-grade systems, industry-specific intermediate models can alleviate some of the challenges. Some of these solutions will be open-source, but many may be bespoke closed source solutions that use audits for quality guarantees.
Key Takeaway: The AI race will shift away from ‘bigger’ models, to more niche domain/task-specific models, and many non BigAI companies will release fine-tuned models that leverage their proprietary data and expertise.
2. It will be an exciting year for multi-modal models
In December, Google unveiled Veo 2 and set a new standard for video-generation models. By the end of this year, we are likely to see several breakthroughs in the multi-modal model space. Key areas to watch are:
Combining video and audio. None of the current video generator models can natively generate audio. In general, models that can take as input a combination of modalities preprocess them into a shared “space”, stream them through a shared network and decode the final output into one modality. The later two steps often use different classes of models for images/video frames and audio outputs. A new approach that can generate the two components simultaneously will represent the next significant step forward.
Longer videos. Google claims that Veo 2 can generate videos up to two minutes in length, but the beta program only allows videos up to 8 seconds in length. Longer videos would require the model to have a memory that tracks parts of the scene that have already been generated. (FYI: Generating long content remains a challenge for text-generation models, as well. Experimental techniques rely on agentic approaches to plan-out the output).
Understanding physics. Research from earlier this year showed that Sora had a limited understanding of real-world physics and that pre-training on 2-D data may not be sufficient to infuse those concepts into the models. Novel post-training techniques that intentionally teach the models about physics may emerge, similar to how language models can be taught to reason mathematically.
These advances will heighten risks associated with generated video content, like deep fakes and non-consensual imagery. To combat them Veo-2 videos are tagged with an invisible watermark and the Sora system uses input and output content filters. While research into new technical mitigations will reduce some of the risks, new frameworks and policy will need to be developed to keep up with the multi-modal developments.
Key Takeaway: Multi-modal models will become more popular in 2025, but they are still limited in some key ways, including their real-world understanding of things like physics, and ability to generate multiple output types at once (such as audio that is aligned with the video).
3. There will be no red or blue state approach to AI
State legislators were busy in 2024 trying to enact a range of AI regulations. The biggest news at the state-level last year was Colorado’s SB 205, the first comprehensive state AI law. As state legislatures reconvene this year, lawmakers in places like Texas and Virginia are considering their own comprehensive state AI laws for high-risk AI systems. We previously examined the Texas legislation and American Legislative Exchange Council’s model AI legislation; noting the commentary about both serving as blueprints for red state lawmakers. Yet, there appears to be no meaningful ideological differences between the bills in Texas and Virginia, as compared to the Colorado law.
Generally, the scope and key obligations across the laws and legislation are similar. For instance, each law or bill requires AI developers and deployers to maintain governance programs for their high-risk AI systems. This includes requiring impact assessments for high-risk AI systems and referencing the NIST AI Risk Management Framework as baseline guidance for organizations to use when managing their AI risks. While the “red state, blue state” divide might work for driving headlines, the AI laws that are being proposed or enacted do not reflect the same ideological divide. Going forward, states will likely take a copycat approach to AI laws rather than one that exhibits meaningful ideological differences.
Key Takeaway: We haven’t really seen a ‘red state’ AI bill with a lot of traction that is differentiated from what ‘blue’ states have pursued. That being said, with slim margins in Congress, and their heavy interest to accelerate AI, we do expect a lot of State-level activity.
4. Leveraging Federal Agency authorities to regulate AI will be difficult under the Trump Administration
It is unlikely that Congress will enact sweeping AI rules in the next couple of years, as Washington prepares for Republicans to be in full control. However, last year the House and Senate released bi-partisan reports that recommend the federal government leverage existing law to help oversee AI. This approach would require federal agencies to use their existing authorities to interpret the appropriate use and oversight of AI systems. Yet, this could directly conflict with President-elect Trump’s goal of reducing the administrative state.
Federal agencies were already hamstrung last year, after a series of Supreme Court decisions limited federal agencies' rulemaking and enforcement authorities. President-elect Trump’s plan to cut the federal workforce and restructure federal agencies could further restrict their ability to tackle AI-related issues. It is unclear how successful the incoming Administration will be in accomplishing its goals to overhaul the federal government. However, even if little progress is made in reducing the size of federal agencies, the political environment will likely chill meaningful AI-related oversight.
Key Takeaway: Many US federal regulatory agencies, such as the FTC, had been active on AI policy issues, partly due to President Biden’s AI Executive Order. We expect federal agencies will be a lot less active under the Trump administration, and any meaningful AI policy will be more centralized.
5. Agentic AI will not be adopted at scale within enterprises… yet
AI’s new obsession for 2025 is AI Agents. Mark Zuckerberg recently predicted that AI Agents would replace most of Meta’s ‘mid level software engineers’, Gartner listed AI Agents as their #1 tech trend for 2025, agentic AI startups (such as Hippocratic (healthcare), Norm.ai (Legal), and Sierra (Consumer AI Agent)) received record funding last year, and ChatGPT now supports basic ‘agentic’ functions. It’s clear that the hype is real, but despite the headlines, we think widespread adoption of AI agents won’t happen in 2025.
Firstly, the definition of an ‘AI Agent’ is not clear, and they range from systems capable of just copying-and-pasting content, to multi-step actions involving a series of ‘reasonings’. Many organizations may simply rebrand their existing AI features as ‘agentic’, but not actually implement their full capabilities. For the more capable agents, there is a minefield of potential security, legal, and ethical risks that are only starting to get discussed. We do expect some experimentation with agents for low risk use cases, but many businesses can still reap huge productivity gains from non-agentic AI use, without bringing on additional risks, and potential employee backlash.
Key Takeaway: While some in the tech sector are well positioned to leverage agentic AI, it’s still too early for many enterprises. Many organizations are still working to understand how to prove AI ROI, and implement appropriate AI risk management. Once governance is implemented, AI can flourish responsibly within the organization.
*********
As always, we welcome your feedback on content and how to improve this newsletter!
AI Responsibly,
- Trustible team