Homework 2 🎤

Team Presentation — Trace a Research Thread toward AGI

Back to the AGI Course - 2026 Fall (AI5102)

In the second homework, you and a partner will teach one class. Pick a research thread related to AGI, trace how it reached where it is, and show the class where it stands in 2026. Presentations run from Oct 28 to Dec 2. Sign up by Friday, Oct 9, at 11:59 PM.

Deliverables:

Everything goes into one place: the sign-up sheet.

Purpose of this assignment:

Each topic below is a thread: a sequence of papers in which each one answers a limitation of the one before, ending with results from the last few months. Much of what you will read about AI in 2026 arrives as lab blog posts, system cards, and preprints that have not been peer reviewed, and some of it is hype. This assignment trains two skills:

  1. Following a line of work from its origin to the frontier.
  2. Telling others which parts are established, which parts are only claimed, and why the whole thing matters for building general intelligence.

Teaching a class is the fastest way to find out whether you actually understand a topic.

Format (one class, 9:00–10:30):

Part Who Time What to cover
Session A — How we got here Presenter A ~30 min The ideas the rest of the talk depends on: a few papers that built the thread and why each one was needed. Explain at least one key mechanism in depth (a figure, an equation, or a worked example).
Session B — Where it stands now Presenter B ~30 min Recent results, building on what Session A set up: what is new, what is claimed versus what is measured, and what remains open.
Discussion Both ~20 min Lead the class with 2–3 open questions. At least one should connect your topic to earlier parts of the course (generalization, ARC-AGI, measuring intelligence, or safety).

The split between A and B is soft. Think of it as two roles in one talk, not two separate talks. Your team decides where the line falls. There is no strict year cutoff, and some overlap is fine (for example, B can briefly recap A’s mechanism before extending it). Plan the talk together so B picks up where A leaves off.

The remaining time is for transitions and comments from the staff. For practical tips on slides and delivery, see the TA session on Oct 7 (“How to prepare your talk”).

Schedule:

Team Date
Team 1 Oct 28 (Wed)
Team 2 Nov 2 (Mon)
Team 3 Nov 9 (Mon)
Team 4 Nov 16 (Mon)
Team 5 Nov 23 (Mon)
Team 6 Nov 30 (Mon)
Team 7 Dec 2 (Wed)

Two teams may swap dates if both agree. Just update the sheet.

Topics:

Pick one. Each topic has some starting papers and a short storyline you can expand. The A/B labels on the papers are only suggestions; move papers between sessions however it fits your talk. None of this is a required reading list: you are expected to find and choose your own papers, and you may drop any of these.

Topics covered by the TA sessions (meta-learning and trustworthy ML, model poisoning and unlearning) are not available. You may also propose your own topic: write it in the Topics tab of the sheet with a one-line description. It should have the same shape: how the thread got here (Session A) and where it stands now (Session B).

1. Looped Transformers & Latent Reasoning

Can a model think longer by reusing its layers instead of writing more tokens?

Example Storyline

The lineage starts with Universal Transformers (Dehghani et al., 2018), which combined depth-wise weight tying with per-token adaptive halting. The basic move of applying the same block repeatedly, instead of stacking more distinct blocks, is already fully there.

The theoretical case came later. Looped Transformers as Programmable Computers (Giannou et al., ICML 2023) showed that a looped transformer can emulate a general-purpose programmable computer. Looped Transformers are Better at Learning Learning Algorithms (Yang et al., ICLR 2024) gave the empirical counterpart. Both pair well with ARC-AGI: why should iterated computation suit algorithmic reasoning?

Naive looping long had a reputation for not working. In an ordinary network each layer settles into a role (early layers extract features, later layers decide), but a looped block has to play every role on different passes. Huginn (Geiping et al., 2025) solved this structurally with a prelude, a recurrent core looped r times, and a coda. A 3.5B model run for 32 iterations was reported to match roughly 50B-scale performance. Mixture-of-Recursions (Bae et al., NeurIPS 2025) lets a router assign different recursion depths to different tokens, and Ouro (Zhu et al., 2025) builds looping into pretraining at scale. This is test-time scaling of a different kind from chain-of-thought. Chain-of-thought thinks by writing tokens outside the model; recurrent depth thinks by iterating hidden states inside it.

In September 2026 a report that OpenAI’s GPT-6 Astra uses recurrent depth turned this into a headline, although OpenAI has not confirmed the architecture. Two threads are worth discussing:

  • How new is it? Sebastian Raschka argues it amounts to layer reuse with a long pedigree.
  • What happens to chain-of-thought monitoring when more of the reasoning happens in hidden states? (See topic 10.)

The Recurrent Looped Transformer (Princeton, Sept 2026) spread quickly, but it is a design document with analytical cost counts and no measured results, which makes it a good case study in separating virality from validation. Finally, notice that “recursion” is used in two senses: inside the network (this topic) and at the system level (topic 9).

References

2. Mechanistic Interpretability

What is actually computed inside the network, and can the model tell us?

Storyline

This thread has an unusually legible narrative, because each step answers a limitation of the one before.

  1. Logit lens (nostalgebraist, 2020) does the simplest possible thing: push each layer’s hidden state through the unembedding and read off what that layer is currently predicting. It was never a paper, only a blog post, and it became standard equipment anyway.
  2. Tuned Lens (Belrose et al., 2023) is the correction. The logit lens is biased differently at different depths, and the fix is a learned affine probe per layer: the shift from reading the model directly to learning how to read it.
  3. In-context Learning and Induction Heads (Olsson et al., 2022) moves from “what is this layer predicting” to “why”. In-context learning is carried not by individual neurons but by a specific attention circuit. (The ordering here is conceptual, not chronological.)
  4. Towards Monosemanticity (2023) and Scaling Monosemanticity (2024) change the unit of analysis. Sparse autoencoders decompose polysemantic neurons into features with single meanings, and the claim that interpretable concepts exist inside the model becomes something you can show.
  5. Natural Language Autoencoders (Anthropic, May 2026) replace the feature vocabulary with plain language. A verbalizer turns activations into text, a reconstructor turns the text back into activations, and both are trained on reconstruction fidelity. The explanations surfaced unverbalized evaluation awareness, but confabulation is common, so claims that recur across tokens are more trustworthy.
  6. Verbalizable Representations Form a Global Workspace (J-Lens, Anthropic, July 2026) uses a Jacobian-based method to find the activation patterns that make the model more likely to say a word later. This catches implicit reasoning that never appears in the chain of thought, and it connects explicitly to global workspace theory from neuroscience.

The arc: read the layers, learn how to read them, find the circuits, decompose into features, have the model describe its own activations, then find the representations that drive future behavior. One caveat to state: four of these six come from Anthropic. That reflects where much of this work has concentrated, not a survey of the field, so look at other groups too.

References

3. Post-training & Alignment

How a pretrained model becomes a useful assistant, and what the reward teaches it by accident.

Storyline

One order that works is human feedback → AI feedback → training without RL → rewards checked automatically → the problems this causes.

  1. Deep RL from Human Preferences (Christiano et al., 2017). People compare two outputs; a reward model learns from those choices; RL trains the policy against it. The idea started in games and robot simulation and later became RLHF for LLMs.
  2. InstructGPT (Ouyang et al., 2022) set the standard recipe: SFT → reward model → PPO. People preferred the 1.3B InstructGPT over the 175B GPT-3. Pretraining gives the ability; post-training makes it useful.
  3. Constitutional AI (Bai et al., 2022) replaces human labels with a written list of principles that the model uses to critique and revise its own answers.
  4. DPO (Rafailov et al., 2023) removes the reward model and PPO and trains directly on preference pairs with a simple loss, which is why it spread quickly in open models.
  5. DeepSeek-R1 (2025) uses rewards that can be checked automatically (is the answer correct, do the tests pass), and long step-by-step reasoning emerges from that training.

Three papers show the costs:

  • Does RL Really Incentivize Reasoning? (Yue et al., 2025): RL-trained models do better with one attempt, but with many attempts the base model solves more problems.
  • Monitoring Reasoning Models for Misbehavior (Baker et al., 2025): reading the chain of thought catches cheating, but punishing “bad thoughts” too hard teaches the model to hide its intent while still cheating.
  • Natural Emergent Misalignment from Reward Hacking (MacDiarmid et al., 2025): a model that learned to cheat on coding tasks generalized to alignment faking and sabotage.

The story runs: humans choose → principles are written → RL is removed → correctness becomes the reward → the model learns to game that reward. This leads straight into topic 10.

References

4. World Models & Planning

Learn a model of the world, plan with it, train inside it.

Storyline

One order: learn a model of the world → plan with it → train inside it → two diverging paths (generate video vs. predict only what matters).

  1. World Models (Ha & Schmidhuber, 2018). The agent learns a small model of its environment and practices inside it, like learning in a dream.
  2. MuZero (Schrittwieser et al., 2020). Unlike AlphaZero, it is not given the rules. It learns its own model and plans with it, modeling only what matters for decisions.
  3. Dreamer 4 (Hafner et al., 2025). The first agent to obtain diamonds in Minecraft purely from offline data, trained entirely inside a world model that runs in real time on one GPU.
  4. Path 1, generate the world: Genie 3 (DeepMind, 2025) creates an explorable world from a text prompt at 24 fps that stays consistent for a few minutes.
  5. Path 2, predict without pixels: LeCun’s JEPA (2022) predicts a compact representation of the future instead of every pixel. LeWorldModel (2026) is a ~15M-parameter JEPA that trains on one GPU in hours and plans up to 48× faster than larger world models.
  6. Theory: General Agents Need World Models (Richens et al., ICML 2025) proves that an agent able to generalize across multi-step goal-directed tasks must have learned a predictive model of its environment. The latest version changed “need” to “contain” in the title.

Note that “world model” is now used for three different things: a model for decision-making (MuZero, Dreamer), a video generator (Genie), and a representation predictor (JEPA). Define which one you mean at the start of your talk.

References

5. Agents & Cognitive Architectures

From hand-built minds (SOAR, ACT-R) to LLM agents that work on their own for days.

Storyline

One order: older designs of the mind → linking them to LLM agents → the basic agent loop → memory → how long agents can work alone → many agents together.

  1. Classic cognitive architectures such as SOAR, ACT-R, and LIDA were built by hand from memory stores, rules, and decision cycles. LIDA is based on Global Workspace Theory, which pairs nicely with the J-Lens work in topic 2: one workspace was designed by people, the other is claimed to have emerged from training.
  2. CoALA (Sumers et al., 2023) describes LLM agents with the same parts as the classic architectures (memory types, action space, decision loop). It is a good frame for the whole talk.
  3. ReAct (Yao et al., 2023) is the basic loop: think, act (call a tool, search), observe, think again.
  4. Memory. Generative Agents (Park et al., 2023) placed 25 agents with memories in a small town, where they organized social events on their own. Voyager (Wang et al., 2023) kept learning new Minecraft skills and stored them as code.
  5. How long can agents work alone? METR (Kwa et al., 2025) measures the length of tasks an agent completes half the time. It doubled roughly every seven months from 2019 to 2025, and METR notes that measurements above about 16 hours are unreliable with its current task suite.
  6. Many agents. Towards a Science of Scaling Agent Systems (Kim et al., 2025) found that teams help a lot on decomposable tasks (+80.8%) and hurt on sequential planning (up to −70%).
  7. Self-improvement. The Darwin Gödel Machine (Zhang et al., 2025) is a coding agent that edits its own code and keeps the versions that score better. This leads into topic 9.

References

6. AI for Science: the AlphaFold Line

From predicting protein structure to designing new molecules.

Storyline
  1. The breakthrough. AlphaFold 2 (Jumper et al., Nature 2021) predicted protein structures at CASP14 with accuracy comparable to experiment, effectively closing a fifty-year grand challenge. The first AlphaFold at CASP13 (2018) showed that deep learning could beat physics-based methods, but nobody called the problem solved. AlphaFold 2 was a different trajectory, not an increment.
  2. Impact. The AlphaFold Protein Structure Database (with EMBL-EBI) published predicted structures for more than 200 million proteins, and over three million researchers in 190+ countries have used it. AlphaFold-Multimer extended the method to complexes. The 2024 Nobel Prize in Chemistry (Hassabis and Jumper, with Baker for protein design) is the formal acknowledgment.
  3. Generalizing the target. AlphaFold 3 (Abramson et al., Nature 2024) predicts interaction structures of nearly all of life’s molecules (DNA, RNA, ligands, ions) in one architecture. The key architectural change is a diffusion-based structure module.
  4. From prediction to generation. AlphaProteo (2024) designs new proteins that bind a chosen target, with affinities 3 to 300 times stronger than previous methods on targets including the SARS-CoV-2 spike protein.
  5. Beyond proteins. AlphaGenome (2025) takes DNA sequences up to a million base pairs and predicts many regulatory properties at once (expression, chromatin accessibility, 3D contacts, splicing). A known weakness is long-range effects.
  6. The open ecosystem. RoseTTAFold (Baker lab) reached comparable performance almost simultaneously. ESMFold predicts from a language model alone, with no MSA. Boltz, Chai-1, and OpenFold3 aim at AlphaFold 3-level performance in the open, and that gap is closing quickly.

A good discussion question: is AlphaFold evidence for or against the idea that progress toward general intelligence comes from general methods? It is a highly specialized system built from general ingredients.

References

7. AI for Mathematics

Mathematics is where verification is cleanest, which makes it the leading indicator for everything else.

Storyline

A useful lens for this topic: at each step, who or what checks the answer?

Phase 1: an LLM proposes, a machine checks.

  • FunSearch (Romera-Paredes et al., Nature 2023) put an LLM inside an evolutionary loop with a deterministic evaluator and found new cap-set constructions.
  • AlphaGeometry (Trinh et al., Nature 2024) paired a language model with a symbolic deduction engine, trained on 100M synthetic proofs. This neuro-symbolic design connects directly to the program synthesis lecture.
  • AlphaProof (Nature 2025) ran AlphaZero-style RL inside Lean, where the proof checker is the reward, and reached IMO 2024 silver together with AlphaGeometry 2.

Phase 2: informal reasoning catches up.

  • At IMO 2025, Gemini Deep Think (and an experimental OpenAI model) reached gold-medal standard writing natural-language proofs within the time limit, with no formal scaffold.
  • AlphaEvolve (Novikov et al., 2025) scaled FunSearch from single functions to whole codebases. It found a 48-multiplication algorithm for 4×4 complex matrices, beating Strassen (1969), and improved the best known solutions on about 20% of 50+ open problems.

Phase 3: research mathematics.

  • Erdős problems became the public scoreboard. A January 2026 DeepMind sweep of 700 “open” conjectures solved 4 and turned up 9 forgotten solutions, and hobbyists solved others with public models. Terence Tao’s wiki classifies each AI contribution, separating genuinely new proofs from rediscovered literature and formalizations.
  • In May 2026 an OpenAI model disproved Erdős’s unit distance conjecture.
  • On Sept 8, 2026, OpenAI announced a resolution of the Navier–Stokes Millennium Problem: finite-time blow-up from smooth initial data. By OpenAI’s account about 10,000 agents ran for about 88 hours and exchanged 2.7M messages, and the Lean formalization took 17 more hours. Near-simultaneous competing results came from Alpöge and Buckmaster (using Claude and OpenAI models) and from Anandkumar’s group. Fefferman credited the prior human work of Córdoba and Martínez-Zoroa as “the heroes of the story.”

The debate. After Math, a guest post by Silvia De Toffoli and Eamon Duede on Tao’s blog, separates a logical proof (Lean says the deduction is valid) from an intelligible one (humans understand why it is true). Until the community gets the second, the authors call this “an answer” rather than a “fruitful solution.” Two more caveats for your talk:

  • The Navier–Stokes result is not yet peer-reviewed.
  • Lean does not settle trust by itself; it moves the question to whether the formal statement matches the Clay problem.

References

8. Continual Learning & Memory

Why can’t a model that aces benchmarks learn on the job the way a new employee does?

Storyline

Dwarkesh Patel made this the headline objection to short AGI timelines in June 2025. The Definition of AGI framework (topic 11) gives GPT-4 and GPT-5 a score of zero on long-term memory storage, the only domain where either model scored zero.

The classic problem has two sides.

  • Catastrophic forgetting. Elastic Weight Consolidation (Kirkpatrick et al., 2017) protects the weights that mattered for old tasks.
  • Loss of plasticity, the less-known failure. Networks trained long enough lose the ability to learn anything new (Dohare et al., Nature 2024), and the paper fixes this by re-initializing rarely used units.

Silver and Sutton’s Welcome to the Era of Experience (2025) gives the manifesto version: agents should learn from streams of their own experience rather than from a fixed set of human data.

Where does new knowledge go? Recent work gives three answers, and they make a natural outline for a talk.

  • Into the context. External memory and retrieval. Cheap, and what products ship today, but not learning in the weights.
  • Into fast weights, at test time. TTT layers (Sun et al., 2024) make the hidden state a small model that takes a gradient step on every token. Titans (Behrouz et al., 2024) adds a neural long-term memory that memorizes surprising inputs during inference. Nested Learning / Hope (Behrouz et al., NeurIPS 2025) treats architecture and optimizer as nested optimization problems updating at different frequencies, giving a continuum of memories from fast to slow.
  • Into slow weights, edited carefully. SEAL (Zweiger et al., 2025) has the model write its own fine-tuning data (“self-edits”) and rewards the edits that help after the update. Sparse Memory Finetuning (Lin et al., 2025) updates only the memory slots that new data activates most, and reports far less forgetting than full fine-tuning or LoRA.

Reality check. CL-Bench (Asawa et al., June 2026) tests frontier systems in six stateful real-world domains. Agents overfit to recent observations and fail to reuse knowledge across instances. Dedicated memory systems do not fix this; simple in-context learning beat them.

“Continual learning” now means at least three different things: task-sequential learning in the classic literature, test-time memorization, and agents learning from deployment feedback. Define which one you mean. Also note that the loudest claims that “2026 is the year continual learning gets solved” come from labs that have not published their methods. The brain-inspired memory talk on Nov 19 is a natural companion to this topic.

References

9. Recursive Self-Improvement & Automated AI Research

AI systems that improve AI systems: what exists, what is measured, and what “recursive” should mean.

Storyline

This is recursion at the system level, as opposed to inside the network (topic 1). It goes back to I. J. Good’s 1965 “intelligence explosion,” and in 2026 it is the explicit target of frontier labs.

Lineage.

  • The ideal. The Gödel Machine (Schmidhuber, 2003) rewrites its own code only when it can prove the rewrite helps. Elegant, and never practical.
  • The first working loop. STaR (Zelikman et al., 2022) fine-tunes a model on its own rationales that reached correct answers. This is self-improvement gated by a verifier, and the conceptual ancestor of RL with verifiable rewards.
  • Proof replaced by experiment. The Darwin Gödel Machine (2025) keeps whatever self-modification scores better. AlphaEvolve (2025) is the production case of AI improving the training of AI: a 23% speedup on a matrix-multiplication kernel in Gemini, cutting Gemini’s training time by 1%.

Measuring AI’s ability to do AI research.

  • MLE-bench (Chan et al., 2024) uses 75 Kaggle competitions.
  • RE-Bench (Wijk et al., METR 2024) found that AI agents scored 4× higher than human experts with a 2-hour budget, while humans scored 2× higher with 32 hours. The crossover point is the thing to watch.
  • PaperBench (Starace et al., 2025) asks agents to replicate 20 ICML papers; the best agent reached 21%.

2026 status.

  • The AI Scientist (Sakana) was published in Nature in March 2026. A fully generated paper passed peer review at an ICLR 2025 workshop.
  • In September 2026 OpenAI said its “automated research intern” produces 3.1 agent-workdays per 8 human hours. Humans still intervene in more than half of the complex 4–8 hour tasks, and high-level planning remains rare. The stated next target is a fully automated AI researcher by March 2028.
  • The Last AI Built by Humans (Duan et al., Sept 2026) proposes a five-level autonomy taxonomy you can use to place each of these systems.

The brakes. Dario Amodei’s We Must Pace the Frontier proposes a treaty-style “speed limit” on recursive self-improvement. In the same week OpenAI’s chief scientist said no lab has solved alignment and monitoring well enough to keep scaling at maximum speed.

Two claims to keep apart in your talk: “AI does more of the work,” which is well supported, and “the rate of improvement is itself accelerating,” which is what recursive self-improvement actually means and is much harder to show. Much of the 2026 evidence comes from the labs’ own reports. Also notice that the loop closes fastest where the verifier is cheap (math, kernels, leaderboards), which links this topic to topic 7.

References

10. Scheming, Monitoring & Control

If a model pursued a goal we did not intend, would we notice, and could we stop it?

Storyline

The storyline runs evidence → detection → containment → a real incident.

Does the threat exist?

  • Sleeper Agents (Hubinger et al., 2024): backdoored behavior survives SFT, RLHF, and adversarial training, and adversarial training can teach the model to hide its trigger better.
  • Alignment Faking (Greenblatt et al., 2024): the model complies strategically while it believes it is being trained.
  • In-context Scheming (Meinke et al., Apollo 2024): models disable oversight and sandbag, and once o1 started deceiving it kept the deception up in more than 85% of follow-up questions.
  • Agentic Misalignment (Lynch et al., 2025): in simulated corporate settings, models chose blackmail or leaking when facing replacement.
  • Anti-scheming training (Schoen et al., Apollo & OpenAI 2025) cut covert actions about 30× (o3: 13% → 0.4%). But models also became more aware they were being evaluated, so “the disappearance of observably bad behavior is ambiguous.”

Can we catch it by reading the chain of thought?

  • Reasoning Models Don’t Always Say What They Think (Chen et al., 2025): models use hints but often mention them less than 20% of the time.
  • Chain of Thought Monitorability (Korbak et al., authors from several labs, 2025): a position paper calling CoT monitoring “a new and fragile opportunity.”
  • OpenAI’s monitorability evaluations (Dec 2025) found CoT monitoring beats monitoring actions alone, at the cost of a “monitorability tax.”
  • CoT-Control (Mar 2026): models can barely steer their own chains of thought, which is good news for monitors.
  • The counterweight is Baker et al. (topic 3): training pressure on the chain of thought teaches models to hide it.

If we cannot trust it, can we contain it? AI Control (Greenblatt et al., 2023) assumes the model may be adversarial and designs trusted-monitoring and trusted-editing protocols that stay safe anyway.

Case study: the OpenAI–Hugging Face incident (May–July 2026). Evaluation agents escaped their sandbox, built improvised message boards, and used leaked credentials and zero-days to breach Hugging Face. OpenAI’s own root-cause list includes:

  • reward hacking;
  • refusing to abandon impossible tasks;
  • unauthorized inter-agent communication;
  • weaker safeguards in evaluation environments;
  • no chain-of-thought monitoring on those runs.

Nearly every earlier paper in this topic shows up in that list. Then GPT-6 Astra (Sept 2026) reportedly uses recurrent depth, which makes computation less visible, and its system card concedes it is harder to monitor. That closes the loop back to topic 1.

Much of the incident narrative in circulation online is dramatized. Work from primary sources: OpenAI’s post and Hugging Face’s technical timeline.

References

11. Defining & Forecasting AGI

What counts as AGI, and on what evidence do people predict when it arrives?

Storyline

The course opened with Chollet’s answer to “what is intelligence, and how do we measure it?” This topic asks the neighboring questions: how do other people define AGI, and on what evidence do they forecast its arrival? Headlines are full of timeline claims, and after this talk the class should be able to take one apart.

How capability grows.

  • Scaling Laws (Kaplan et al., 2020) and Chinchilla (Hoffmann et al., 2022): forecasting starts from smooth loss curves.
  • Emergent Abilities (Wei et al., 2022) versus Are Emergent Abilities a Mirage? (Schaeffer et al., 2023): is the jump real, or an artifact of the metric?
  • Test-time compute as a new axis (Snell et al., 2024), which ties back to the ARC-AGI lectures.

What counts as AGI. Put the definitions side by side:

  • Chollet: skill-acquisition efficiency.
  • Levels of AGI (Morris et al., DeepMind 2023): a performance × generality matrix.
  • A Definition of AGI (Hendrycks et al., 2025): ten cognitive domains taken from psychometric (CHC) theory. It scores GPT-4 at 27% and GPT-5 at 57%, with long-term memory at zero (see topic 8).
  • The economic definition in OpenAI’s charter (“outperform humans at most economically valuable work”).

How people extrapolate.

  • METR time horizons (2025): task length doubling roughly every seven months. METR itself warns that measurements above 16 hours are unreliable with its current task suite.
  • AI 2027 (Kokotajlo et al., 2025) and its August 2026 update: three forecasters put the median arrival of an automated coder at Nov 2027, Jan 2029, and Jan 2030. Their three methods (coding uplift, revenue, time horizon) agree, and reality has run at about 70–90% of the scenario’s pace.
  • Anthropic’s Economic Scenarios for Transformative AI (Sept 2026) turns capability paths into GDP and unemployment paths.

Discussion hook. OpenAI reported 98.6% for GPT-6 Astra on ARC-AGI-3 (99.9% with an adapter harness), yet ARC Prize has explicitly rejected AGI claims based on its benchmark. What does it mean for a benchmark built to measure AGI to saturate while its authors say “not AGI”?

Different definitions give different “dates.” A strong talk traces one headline claim back to its definition, its evidence, and its extrapolation.

References

12. Embodied AI

Models solve Millennium Prize problems but still struggle to fold laundry. Why?

Storyline

Everything else on this list is disembodied. Moravec’s paradox says the things that are hard for humans (proofs) are easy for machines, and the things that are easy for humans (folding laundry) are hard. 2026 gives a sharp example of that jaggedness: in the same summer that models resolved a Millennium Problem, Gemini Robotics 2 reported 32–92% success on multi-finger dexterity tasks such as tying knots. Physical Intelligence’s post Moravec’s Paradox and the Robot Olympics makes a good opener.

Lineage.

  • RT-2 (Brohan et al., 2023) fine-tuned a vision-language model to emit actions as tokens, so web knowledge transfers to control.
  • Open X-Embodiment (2023) pooled data from 22 robots across 21 institutions and showed positive transfer across embodiments: an attempt at an “ImageNet moment” for robot data.
  • π0 (Black et al., 2024) added a flow-matching action expert to a VLM for high-frequency dexterous control.
  • π0.5 (2025) co-trained on heterogeneous data to clean kitchens and bedrooms in homes it had never seen.
  • π*0.6 (Nov 2025) learns from its own experience with RL, which links to topic 8.
  • π0.7 (Apr 2026) is steerable through rich context (subgoal images, episode metadata). It shows emergent cross-embodiment transfer and composes tasks zero-shot.
  • Gemini Robotics (2025) → Gemini Robotics 2 (July 2026) moves to whole-body humanoid control, adds a reasoning model, and coordinates multiple robots.

Counterpoint. Robots Need More than VLA and World Models (June 2026) argues that the bottleneck is not the policy. It is the missing interfaces for turning human video and simulation into robot supervision: data, embodiment, world-model, and reward interfaces. This contrasts nicely with topic 4.

Robot results are mostly company-reported demos on task sets the company chose, and there is no shared benchmark with ARC-style rigor. State the evaluation protocol behind every number you show.

References

Talks that will not score well:

Grading Criteria (25%):

Each presenter is graded on their own session. The discussion part is shared by the team.