In this year's roundup, Nic, Evertjan, and Cascha unpack the rollercoaster that 2025 was for AI. From the shockwave of DeepSeek R1 to the "gradual" progress of GPT-5 and Gemini 3.0, we break down the tech, the data, (and the drama). 

We discuss why 2025 became the "Year of the Idea Guy" rather than the "Year of General Agents," analyze the Great Minds of AI (from Ilya’s return to research to Dario’s scaling maximalism), and look at what the betting markets say about the bubble bursting. Finally, we make our own predictions on what 2026 holds for the industry.

We cover:

The 2025 Timeline: A month-by-month breakdown of model releases (DeepSeek, o3, Claude 3.7, Gemini 3).

The Pelican Benchmark: Visualizing progress through Simon Willison’s famous "Pelican on a bicycle" test.

Highlights & Lowlights: The rise of agentic coding, the failure of general agents, and the "Information Obscurity" problem in AI labs.

Great Minds in AI: A deep dive into the conflicting philosophies of Sutskever, Karpathy, LeCun, Sutton, Hassabis, and Amodei.

Polymarkets & Predictions: What the crowd thinks about AGI, Tesla FSD, and the AI bubble, plus our team's predictions for 2026.

Links & Resources Mentioned:

Benchmarks & Data:

⁠METR Time-Horizon Benchmark⁠ - Tracking the jump in capabilities.

⁠Epoch AI Capabilities Index⁠ - Aggregated model progress

⁠The Pelican Benchmark (Simon Willison)⁠ - The evolution of SVGs.

⁠Epoch AI Data Center Sizes⁠ - Visualizing the gigawatt scale.

Interviews & Articles:

⁠Ilya Sutskever on Dwarkesh Podcast⁠ - "The Return to Discovery".

⁠Andrej Karpathy on Dwarkesh Podcast⁠ - The "Decade of Agents" and "Silent Collapse".

⁠Sam Altman: The Gentle Singularity⁠ - Referenced in discussion.

⁠Horses vs. Cars Analogy⁠ - Understanding adoption lag.

Prediction Markets (Polymarket & Metaculus):

⁠Will the AI Bubble Burst?⁠ 

⁠Tesla Unsupervised FSD Timeline⁠ 

⁠When will weakly general AI be announced?⁠ 

How can AI support global food security in practice? In this episode, we speak with Marieke Meeske PhD researcher at Tilburg University’s Zero Hunger Lab, about how Natural Language Processing (NLP) is applied to food security analysis and policymaking.

We cover her review of 60+ NLP use cases, from early warning systems to understanding public perceptions around food sustainability. We also explore why most AI solutions in this space struggle to move beyond the research stage, and what this “implementation gap” means for real-world impact.

Finally, Marieke shares insights from her current work with the UN’s FAO, including the use of NLP to measure policy integration across climate, agriculture, and nutrition — helping identify gaps and misalignments in national food policies.

In what ways is the rise of generative AI impacting journalism? 

In this episode, we sit down with Andrii Degeler, founder of unzip.media and journalist-in-residence at Dealroom.co, to discuss what the AI shift means for newsroom workflows, the role of the journalist, and the evolving media business model. We also touch on the overarching themes of media trust and AI literacy.

We unpack how newsrooms are experimenting with GenAI, the ethical questions around authorship and disclosure, the impact of Google’s AI overviews on publishers, and the sustainability of media business models in an AI driven-landscape. 

Links and references:

The new media literacy

Generative AI and news reports 2025: How people think about AI's role in journalism and society

Business Insider reportedly tells journalists they can use AI to draft stories

The Atlantic's AI strategy, explained by its CEO

Global Risks Report 2025 (World Economic Forum)

AI’s role in education is growing fast, but what does it really mean for students and teachers? In this episode, we interview Damian Ashworth, teacher and co-founder of BrainBite – a platform built around hyper-personalization and AI tutors that adapt content to each student’s level, pace, and interests.

We dive into the benefits of AI tutors, the challenges and limitations still ahead, and what a hybrid future of human teachers and AI support might look like.

Along the way, Damian reflects on what AI can (and can not) do for education, and offers advice for both educators and edtech startups.

In this episode, we take a deep dive into the Vesuvius Challenge, which combines ancient history with AI to resurrect an ancient library that was buried in the ashes of Mount Vesuvius in 79 AD. Thought to be lost forever, these scrolls were completely turned to charcoal and are impossible to unroll without destroying them.

We focus on how a team of three university students used a combination of machine learning, computer vision, and X-ray tomography to win the Grand Prize in 2023. We unpack this technical breakthrough, from using an optical flow segmentation algorithm to virtually unroll the scrolls, to training a TimeSformer model to detect the invisible, carbon-based ink.

We also uncover the historical significance of the newly deciphered texts, which belong to the Epicurean philosopher Philodemus of Gadara. It's a journey into a world where AI is not only recovering a lost library but also creating a new methodology — a sort of "Rosetta Stone"— that could be used to decipher other ancient, fragile documents.

The Vesuvius Challenge is ongoing, and there are still prizes to be won! 

Links and references:

Grand Prize 2023: https://scrollprize.org/grandprize

Fragments: https://scrollprize.org/data_fragments

Segments: https://scrollprize.org/data_segments

Scrolls: https://scrollprize.org/data_scrolls

Open prizes: https://scrollprize.org/prizes 

Youseff's explanation of his winning ink detection setup: https://youssefnader.com/2024/02/06/the-ink-detection-journey-of-the-vesuvius-challenge/

In this episode of the KickstartAI Podcast, guest hosts Nils and Ioanna dive into the EU’s landmark AI Act — and the questions it raises about the future of innovation in Europe.

Will the regulation enable safe, responsible progress, or introduce complexity, uncertainty, and a competitive disadvantage for European companies?

They explore how the Act may impact AI development in Europe and challenge efforts to build AI sovereignty. They also discuss the broader need for regulation — and what a more supportive, agile framework could look like.

In this episode, we're joined by philosopher and product architect Paul Kuyer from AxonIQ to dissect the growing concerns around AI safety: from existential risks to the immediate dangers we face today. Our conversation revolves around the “AI 2027” paper, which speculates a plausible path to catastrophe driven by artificial general intelligence (AGI). We dive into two main concerns: 

Misalignment: AI’s goals diverging from human values, potentially leading to harmful sub-goals like self-preservation.

Misuse: Bad actors using AI for malicious purposes, including cyber warfare and social manipulation.

We also discuss the path to AGI, skepticism about AI’s future capabilities, and the distinction between the human brain and neural networks. Lastly, we explore our inclination to anthropomorphize AI, and we address AI’s psychological impact, examining its sycophantic tendencies and potential for misuse in manipulating people.

Links and references:

AI 2027: https://ai-2027.com/

Transformer Circuits (Anthropic): https://transformer-circuits.pub/ 

A Disneyland without children: https://www.lesswrong.com/posts/pk9mofif2jWbc6Tv3/fiction-a-disneyland-without-children

Ilya Sutkever on the AI transformation underway: https://www.youtube.com/watch?v=UL7qQ0E2hJ0

This KickstartAI Podcast episode features Johan de Vries from Dexter Energy, discussing how data flows to manage Europe's complex energy grid, which he calls “Europe's largest machine.” Topics include predictive analytics, market dynamics, and future energy trends.

Highlights

Dexter Energy at a Glance:

• Forecasting: Using weather data to predict renewable energy production at high resolutions.

• Trading optimization: Helping clients strategically participate in energy markets.

• Asset steering: Managing assets like renewable curtailment and battery usage in real-time.

In this episode of the Kickstart AI podcast, hosts Evertjan and Cascha explore how the gap between simulated environments and the real world is being bridged for robotics applications.

The Challenge: Sim-to-Real Transfer

• Moving robots from simulation to real world has been a persistent challenge

• Real-world environments are "messy" with imperfect conditions and unexpected obstacles

• Real-world reinforcement learning is expensive, time-consuming, and risky

Three Key Developments

1. Better Simulators

• Focus shifting from visual realism to predictivity

• Modern simulators achieve much stronger correlation between simulation and real-world performance

• Previous agents exploited simulator flaws rather than learning transferable skills

• Self-supervised learning methods (like DINO and CLIP) improve generalization

• Models learn to focus on important features rather than background context

• Better transfer to real-world conditions with different visual characteristics

• GPU-based simulation provides orders of magnitude speedup

• Neural networks learning world physics directly from video data

• Example: Oasis project learning Minecraft physics from gameplay

• NVIDIA's Cosmos generating diverse training environments from limited data

Applications and Future

• Smart city applications including delivery robots and autonomous vehicles

• Neural networks developing internal "world models" similar to human dreams

• Accelerating progress toward real-world deployment of advanced robotic systems

Links and References

• Sim2Real Predictivity: Does Evaluation in Simulation Predict Real-World Performance?

• Cheating by Sliding in Simulation

• DINO paper, and CLIP paper (self-supervised learning)

• ImageNet Moment for Reinforcement Learning?

• Oasis project (minecraft in a diffusion model)

• NVIDIA's Cosmos model demo
  

In this episode of the Kickstart AI podcast, we explore Vision Language Action (VLA) models - the breakthrough technology powering the latest generation of AI robots.

The Figure AI Helix Breakthrough: