🌁#80: What's in 2025? From Elad Gil, François Chollet, Maxime Labonne, swyx and others

Main topic – a pivotal stage

AI is at a pivotal stage. On one hand, promises and flashy demos are evolving into tangible, impactful innovations – autonomous workflows, specialized systems, tools reshaping industries, and, of course, agents, agents, agents. The groundwork is laid, and we’re entering the building phase. It's also a moment to monetize these breakthroughs while grappling with the broader ethical questions surrounding AI.

On the other hand, AI remains inherently unpredictable. Fundamental debates persist –what does AGI truly mean? Are agents genuinely functional? And how are scaling laws shifting under the weight of progress?

We’ve collected a few incredibly interesting predictions for you from François Chollet, Elad Gil, Maxime Labonne and other top experts in the field →

One thing is absolutely clear: there will be no AI winter in 2025. This field is hot.

The year of time-inference search (and AGI)

François Chollet, the mind behind Keras and ARC-AGI project, predicts a transformative shift in AI capabilities driven by inference-time search. On Twitter, he shared, “Inference time search is going to be the big driver of AI capabilities over the next few years. It will simultaneously make AI a lot more useful, driving demand, and make serving more expensive, increasing costs (and prices). Inference hardware makers (NVDA, AMD) will benefit.” Speaking exclusively to Turing Post, he elaborated: “It's going to be the year of inference-time search, and we will see many efforts to create an open source reproduction of O1. Also ARC-AGI will be solved.”

John K. Thompson, Global AI Leader in EY and an author of 4 books on AI, Data, & Analytic Teams (his new book coming next year!) remains skeptical about AGI: “AGI does not arrive in 2025; more like 2250 in reality. I am not talking about glimmers of AGI; I am talking about real intelligence in a non-human form. This accomplishment will be game-changing, and it will take another 100 to 200 years.”

Clearly, in 2025 we continue to argue what AGI means. 

swyx, who coined the term “AI engineer” and made it fancy, takes it to more practical level: “Flow engineering/"guided chain of thought" will be deployed much more in production than o1/reasoning type models.” 

We think that in 2025, AI will prioritize "thinking before speaking," placing inference at the core of machine intelligence. This shift towards test-time compute will spotlight symbolic reasoning, adaptive scaling, and smaller models capable of exceptional problem-solving. Size supremacy will lose its appeal, challenging Nvidia’s hardware dominance and sparking a marketplace of innovation. A smarter, leaner AI era is on the horizon.

The year of smaller models

Smaller models are capturing the imagination of many. Microsoft Research’s Ronen Eldan (co-author of the Phi model family) foresees 2025 as pivotal for AI-assisted mathematics. He told Turing Post, “I believe 2025 will be the first year when language models will begin to assist mathematicians in a significant way. These models will be able to generate proofs for some self-contained lemmas and intermediate results in a paper, given minimal guidance. While they won’t work for all such steps, they’ll handle enough to make a meaningful contribution beyond just improving writing or style.”

Will Schenk, AI practitioner and founder of Focus.AI, echoes this enthusiasm for smaller, targeted models, emphasizing their practicality: “Smaller, tighter models. Advances in training to focus on reasoning with less surplus knowledge. People will start to get a feel for how smaller models can be helpful, and how they can be used in smaller, targeted ways. Cost-effective and possibly on-device.” 

Maxime Labonne, Head of Post-Training at LiquidAI, predicts that “Edge LLMs will go from feasible to popular, thanks to progress with small language models and optimized inference.”

The year of multimodality and new professions

Maxime Labonne also highlights that “Multimodality will become a standard for new models in (mid) 2025. We'll see more audio too.”

swyx, being more practical again, says: “Every chatbot input will also have realtime voice! in the year 2025 it will be inconceivable that every text input box does not always come with an instant-start voice input option.“

Will Schenk adds that “True multimodal models will appear that will let you interact and manipulate video over longer timeframes.”

We think that further advances in language and multimodality will lead to Communication Over Prompting: Prompting as we know it is dying, replaced by a new emphasis on seamless communication. Just as people once learned to navigate Google, by 2025, everyone engaging with AI will instinctively know how to communicate with it – via text, audio or images. This evolution reduces the need for "prompt engineers" and gives rise to broader roles like AI Managers and AI Curators – professionals who guide and refine AI’s integration into workflows, ensuring it remains accurate, ethical, and effective. The Chief AI Officer will gain even more prominence, leading strategic AI initiatives in businesses and institutions.

The year of agents

John K. Thompson highlights that 2025 “will be the year of AI Agents, no surprise there. In 2025, we will see millions of agents deployed. Simple Agents, Intelligent Agents and near the end of 2025 we will start to see the beginnings of agents building agents with the arrival of Polymorphic Agents.”

It’s almost a universal consensus that Agents and Agentic Workflows will continue capturing everybody’s attention in 2025. What we’ve noticed from our Agentic Workflow Series is that people struggle with misconceptions about what “Agent” even means. This post with vocabulary around Agentic Workflows unexpectedly got 425 likes, proving the frustration. We think it will not be fixed in 2025.

The year of changes and disruptions

Ollie Carmichael, Product Manager at Adarga, thinks that “Healthcare and specifically personalized medicine will experience the most disruption. As public health providers (e.g. the NHS in the UK) continue to be plagued by rising costs and patient backlogs, there will be a drop in the cost and practical barriers to entry of bespoke, private personalized medical services that integrate genetic, lifestyle, and clinical data to tailor treatment plans for each individual. The integration of real-time feedback features will move AI-driven personalised healthcare into the mainstream.

Maxime Labonne brings it to the companies level “OpenAI will lose its leadership with increased competition from Anthropic and Google in terms of frontier models.”

We think that Google will start to dominate the scene.

The year of challenges (and open-source)

Maxime Labonne warns that “The open-source landscape will massively change, with established companies switching their strategy toward less permissive licenses or different tiers of models.”

Ollie Carmichael warns that “Perceptions of a US/China AI rivalry are going to reach new levels in 2025. As a result, state-of-the-art AI research will become increasingly interlinked with national security considerations. The realist reaction to the opportunities and risks afforded by cutting edge research will cause friction amongst a US research community that is ostensibly international and 'liberal' in composition. As a result, there will be a split in the community. Proprietary/open-source model dividing lines will start to emerge with those rejecting the national security consensus falling into the latter camp, whilst those who embrace it falling into the former. Equivalents to Google's 2018 Project Maven scandal will emerge and investment cycles will be affected.”

swyx notifies “There will be a major court ruling on the media-AI lawsuits and it will be in favor of AI labs training on public content, but with restrictions on exact reproductions.”

We think that the AI gold rush is giving way to a wave of consolidations. By 2025, we’ll see the strongest players buying up smaller innovators and absorbing their technologies, while weaker companies fall by the wayside. This consolidation will streamline innovation but might also stifle diversity and competition. The question is: will these larger entities pave the way for broader adoption, or will they create walled gardens that limit access to transformative AI technologies? The role of open-source becomes even more important.

From the perspective of Investor

As we said in the beginning, 2025 is also the year for AI to start making money. But it will still require a lot of money from investors. That’s why the opinion of Elad Gil, Silicon Valley powerhouse, investor in Stripe and Airbnb, ex-Twitter VP, author of High Growth Handbook, was so interesting to us.

These are his predictions:

We back Elad on the uncertainty. It will be a turbulent ride!

Twitter library

Last week was also all about NeurIPS. We made this collection for you, it’s a must read:

We are reading/watching

• Beyond Fairness in Computer Vision: A Holistic Approach to Mitigating Harms and Fostering Community-Rooted Computer Vision Research - a detailed report by Timnit Gebru and Remi Denton. An important read.

• Elon Musk wanted an OpenAI for-profit – OpenAI publishes more emails with Elon

• Why CEO Sundar Pichai never panicked - Interview by Semafor

News from The Usual Suspects ©

Today we concentrate solely on the incredibly impressive models that have been launched last week. The AI world has no mercy and knows no holiday chill:

• Google’ Steals the Spotlight

  • Google’s Veo 2 revolutionizes AI video generation, delivering 8-second, 4K cinematic clips that outshine competitors like OpenAI's Sora. With superior physics, clarity, and camera control, Veo 2 sets a new standard for creative storytelling. Available in Google Labs’ VideoFX, this launch signals a leap forward in AI-driven media production.

  • Google’s Gemini 2.0 delivers "agentic" AI, blending multimodal prowess with tool integration for smarter, action-ready systems. With a focus on safety and responsibility, Gemini 2.0 defines the future of intelligent, action-oriented AI.

  • Not a model but a good initiative: Google’s Gemini API opens new doors for researchers, offering multimodal fine-tuning and a staggering 2M-token context window. From embodied AI to sustainability, Gemini accelerates breakthroughs while driving real-world impact. Applications for academic credits via the Gemini Academic Program are now live.

• Cohere’s cool updates

  • Cohere for AI introduces Maya, a multilingual vision-language model bridging cultural gaps in AI. With a 4.4M-sample dataset, toxicity-free refinement, and superior VQA performance, Maya excels across languages while addressing ethical concerns. The model sets a strong benchmark for inclusive AI.

  • Cohere’s Command R7B proves small can be powerful. Designed for edge devices, it balances speed and performance, enabling cost-effective AI democratization without compromising quality.

• LG’s EXAONE 3.5: AI for the Real World

  • LG AI Research debuts EXAONE 3.5, a suite of instruction-tuned LLMs excelling in real-world tasks and long-context comprehension. With top results on 11 benchmarks and efficient scaling, EXAONE 3.5 offers cutting-edge capabilities on Hugging Face for non-commercial research.

• Phi-4: Microsoft’s Logical Marvel

  • Microsoft’s Phi-4 compact SLM tackles advanced reasoning tasks with ease. Built for Azure, it blends precision and efficiency, proving small models can deliver big impact responsibly. We broke this news here. Follow us on Twitter to be the first to know.

There will be no other research papers for this week. Though we prepared the collection, it will be too long of a newsletter to post it here. We will provide a two-week collection next week. With recommended NeurIPS papers and amazing models’ papers – there is enough reading for a week.

Thank you for your support and for reading Turing Post. We appreciate you.