# The Other Bets: Neurosymbolic, World Models, Brains and Evolution > Cognitive architectures, neurosymbolic AI, world models, brain emulation, Thousand Brains, active inference, evolution and AIXI: the other roads to AGI. - Author: Tai Pham Phu - Published: 21 Oct 2026 - Tags: AGI, AI History, Applied AI, Neurosymbolic AI - URL: https://phamphutai.com/2026/10/21/the-other-bets-on-agi/ In July 2024, an AI system sat the International Mathematical Olympiad, the hardest maths competition for high-school students in the world, and scored 28 out of 42. That's a silver medal. The gold line that year was 29. The headlines were about the score. The interesting part was **how** it got there. Google DeepMind didn't use one model. It used two very different kinds of machine bolted together. **AlphaProof** paired a language model with the game-playing reinforcement learning of AlphaZero, and it wrote its answers in **Lean**, a formal language where every step of a proof is checked by a computer. If the checker says the proof is valid, it's valid. No hand-waving. **AlphaGeometry 2** handled the geometry problem, pairing a Gemini-based language model with a symbolic engine that grinds through geometric deductions. It solved that problem in 19 seconds once it had the formal version. There were catches, and DeepMind was open about them. Humans translated the six problems into formal maths before the machine started. Students get two sessions of 4.5 hours. The system solved one problem in minutes but took **up to three days** on others. And it solved none of the two combinatorics problems. Still, it showed that the old dream from Part 2, logic and rules, wasn't dead. It had found a partner. Parts 2 to 6 covered the five big bets: rules, learning, reward, scale, and reasoning. This part covers the rest, the bets that never became the main road but never went away. For each: **the idea, the best result, the wall, and the status in October 2026.** --- ## A map of the other roads There are a lot of them, so here's a map first. Two questions sort them: - **Symbolic or learned?** Is knowledge written down as explicit rules and symbols, or learned from data as numbers? - **Brain-inspired or engineered?** Does the design try to copy how brains work, or does it just try to work?
THE OTHER BETS, ON ONE MAP symbolic learned brain-inspired engineered SOAR / ACT-R models of the mind active inference free energy Thousand Brains cortical columns brain emulation copy the wiring robots that grow embodied learning JEPA world models predict summaries evolution and open-endedness neurosymbolic AlphaProof, AlphaGeometry AIXI the maths ideal today's LLMs
A rough map, not a precise one: where each bet sits between symbols and learning, and between copying the brain and just engineering something that works.
Today's language models, the dashed box, sit in the bottom-right corner: learned, and engineered with little attempt to copy the brain. Almost every bet in this post is, in some way, a reaction to that corner. --- ## 1. Cognitive architectures: build the whole mind **The idea.** Most AI research builds one ability at a time. A cognitive architecture tries to build **the whole mind at once**: memory, attention, learning, decision-making, all in one fixed design that every task runs on. The inspiration was Allen Newell's call for "unified theories of cognition". **Who and when.** **Soar** began at Carnegie Mellon in the early 1980s with John Laird, Allen Newell and Paul Rosenbloom. Its core idea is that all thinking is search through a "problem space", and that when the system works out how to get past an impasse, it saves the shortcut as a new rule. That learning trick is called **chunking**. **ACT-R**, from John Anderson's group at Carnegie Mellon, grew out of his ACT theory from the 1970s. It is focused on human memory and aims to predict how real people perform, down to their reaction times. **Best result.** Two, in very different worlds. - **TacAir-Soar** flew simulated military aircraft. In a 1997 exercise called STOW-97, it ran for **48 continuous hours and flew all the US fixed-wing aircraft** in the simulation: taking orders, coordinating with humans and other simulated pilots, and changing plans mid-flight. - **Cognitive Tutor**, an algebra tutor built on ACT-R models of how students solve problems, went into US schools through Carnegie Learning. A large randomised trial run by RAND, covering over 17,000 students in 147 schools, found **no effect in the first year and a gain of about a fifth of a standard deviation in the second.** Modest, but real, and in a field where most edtech shows nothing. **The wall.** The same wall as symbolic AI in Part 2. Every new domain needs people to write the knowledge in by hand. The architecture is general; the knowledge isn't. And nothing in these systems learned from raw pixels or raw text at scale. **Status in 2026.** Soar is maintained by Laird's group at the University of Michigan, and ACT-R remains a standard tool in cognitive psychology. Their bigger legacy is the vocabulary. When people design an **AI agent** today with "working memory", "long-term memory", a planning loop and tool use, they're drawing a cognitive architecture, just with a language model in the middle. (More on that in [what an AI agent actually is](https://phamphutai.com/2026/10/14/what-an-ai-agent-actually-is/).) --- ## 2. Neurosymbolic AI: learning plus logic **The idea.** Neural networks are good at intuition: spotting patterns, guessing what's promising. Symbolic systems are good at rigour: following rules exactly and checking every step. Neurosymbolic AI tries to get both, with the network proposing and the symbolic part verifying. **Best result.** **AlphaGeometry**, published in *Nature* in January 2024, solved **25 of 30** olympiad geometry problems within the time limit. The previous best automated method solved 10. The average human gold medallist solved 25.9 of the same problems. The division of labour is neat. A symbolic engine derives every fact it can from the diagram. When it gets stuck, the language model suggests one new construction, like "add a point here" or "draw this circle". That opens new paths, and the engine grinds again. The model trained on **100 million synthetic examples**, with no human demonstrations. Its successor, **AlphaGeometry 2**, reported an **84%** solve rate on all IMO geometry problems from 2000 to 2024, up from 54% for the first version. Gary Marcus, a long-time champion of neurosymbolic AI, called it impressive but pointed out a catch: those problems were translated into the system's formal language by hand. Starting from plain English, he estimated, the rate would be much lower. **The wall.** Two walls. First, **translation**: the symbolic half only works once someone has turned the messy problem into precise formal language, often the hardest step. Second, **coverage**: formal tools exist for geometry and some maths and software, not for "should I hire this person". **Status in 2026.** Here's the twist. One year after AlphaProof's silver, in July 2025, an advanced version of **Gemini Deep Think** scored **35 out of 42** at the IMO, a gold-medal score, officially graded by IMO coordinators. It worked **entirely in natural language**, within the 4.5-hour time limit, with no Lean and no hand translation. OpenAI also announced a gold-level result on the same problems, self-reported at the time. So the pure language-model road overtook the neurosymbolic one at its own showcase. But the neurosymbolic idea didn't lose so much as get absorbed. AlphaProof's method was published in *Nature* in November 2025, and formal verification remains the only way to be *certain* a long proof is right. The likely future isn't "symbols versus networks". It's networks that write, and checkers that verify. --- ## 3. World models and JEPA: LeCun's alternative **The idea.** Yann LeCun, Turing Award winner and Meta's long-time chief AI scientist, argues that today's language models are on the wrong road to human-level intelligence. His objections, in short: - Generating one token at a time, with no way to plan the whole answer, lets small errors pile up. - Text is a thin slice of what intelligence needs. A child learns a huge amount about the physical world by watching, long before reading. - Predicting every pixel or every word is wasteful. A good model should predict an **abstract summary** of what happens next and ignore the details that don't matter. His proposed design is **JEPA**, the Joint Embedding Predictive Architecture, built around a learned **world model**: an internal simulator that predicts what your actions will do. I covered this debate, with evidence on both sides, in [World Models: Can AI Simulate Reality?](https://phamphutai.com/2026/10/20/world-models/). The short version: **Best result.** Meta's **V-JEPA 2** (June 2025) learned from over a million hours of video, then needed under 62 hours of robot video to let robot arms in new labs pick up and place objects with no task-specific training. **The wall.** It's early. No JEPA-style system has yet matched the broad, general usefulness of language models. The bet is that it will matter most for robots and the physical world, where LLMs are weakest. **Status in 2026.** LeCun left Meta in late 2025 to start AMI Labs, a company built around this idea. It's one of the best-funded alternative bets in AI right now, and one of the few run by someone who helped invent the dominant approach. --- ## 4. Whole brain emulation: just copy it **The idea.** Forget understanding intelligence. Scan a brain in enough detail, every neuron and every connection, and run it as a simulation. If the copy is faithful enough, it should think like the original. **Who and when.** The founding document is **Whole Brain Emulation: A Roadmap** by Anders Sandberg and Nick Bostrom, a 2008 technical report from Oxford's Future of Humanity Institute. It came out of a 2007 workshop and set out what you would need: scanning technology, a way to turn scans into wiring maps, models of how each neuron behaves, and enough computing power to run it all. It also defined levels of success, from a copy that just behaves like the species, all the way to one that preserves an individual. **Best result.** The wiring maps, called **connectomes**, have come a long way. - In October 2024, the **FlyWire** consortium published in *Nature* the first complete wiring diagram of an adult fruit fly brain: **139,255 neurons** and about **50 million** chemical synapses, mapped from electron microscope images by a mix of AI and human proofreaders. - A companion paper by Philip Shiu and colleagues built a simple simulation on top of that map. Activating the simulated sugar-sensing neurons **correctly predicted** which neurons drive feeding, and the team confirmed predictions in real flies. A wiring diagram alone, with simple neuron models, already told biologists something new. - In April 2025, the **MICrONS** project published a map of one cubic millimetre of mouse visual cortex, about the size of a grain of sand: roughly **200,000 cells and over 500 million connections.** **The wall.** Scale and detail. A human brain has around 86 billion neurons, over 600,000 times more than the fly. And a wiring diagram is not a working brain. The fly simulation had no learning and none of the chemical signals that change how neurons behave. Nobody knows how much of that detail an emulation needs. **Status in 2026.** Real neuroscience, not science fiction. But as a road to AGI it's the slowest on this list. --- ## 5. Thousand Brains: copy the cortex's design **The idea.** Jeff Hawkins, who founded Palm Computing before turning to neuroscience, argues that the neocortex, the wrinkled outer layer of the brain, is built from about 150,000 repeating units called **cortical columns**. In his 2021 book *A Thousand Brains*, he proposes that **each column learns a complete model of objects**, using **reference frames**, a kind of internal map of where things are relative to each other. The columns then **vote** to agree on what you're sensing. And crucially, learning comes from **movement**: you learn what a cup is by moving your fingers and eyes over it. **Best result.** In November 2024, his company Numenta launched the **Thousand Brains Project**, open-source and partly funded by the Gates Foundation. It became an independent non-profit in January 2025. Its first system is called **Monty**, after Vernon Mountcastle, the neuroscientist who proposed cortical columns as the brain's repeating unit. A July 2025 paper reported that Monty can recognise 3D objects and their orientation by moving a sensor over them, learning quickly and continuously, with far less computation than a comparable Vision Transformer. **The wall.** It's tiny compared to the mainstream. The tests are on household objects, not on language, maths or real-world tasks. The authors themselves call the system "nascent". And the theory, while influential, is still debated among neuroscientists. **Status in 2026.** An open research project, explicitly described as not production-ready. Worth watching because it targets exactly what LLMs do badly: **learning continuously, from little data, through a body**. --- ## 6. Active inference: the free energy principle **The idea.** Karl Friston, a neuroscientist at University College London, proposed in a famous 2010 paper, *The free-energy principle: a unified brain theory?*, that brains do one thing: **minimise surprise**. More precisely, they minimise a quantity called "free energy", which is an upper bound on how surprised the brain is by its senses. You can reduce surprise in two ways. Update your beliefs to match the world (that's perception and learning). Or act to make the world match your beliefs (that's action). **Active inference** is the version where action is just another way of making your predictions come true. Instead of chasing rewards, as in Part 4, the agent pursues the outcomes it expects and prefers. Notice the question mark in the paper's title. Friston himself wrote that the principle's ability to unify brain theories was yet to be established. **Best result.** As a theory of the brain, it's one of the most cited ideas in modern neuroscience. As AI, results are small. The most notable recent claim comes from a company, **VERSES AI**, which reported in June 2025 that its active-inference model AXIOM beat DeepMind's DreamerV3 on a set of ten simple arcade games, learning from far fewer interactions with a far smaller model. That's a company claim on a benchmark the same team designed, and I haven't found an independent replication. **The wall.** Critics argue that the principle is so general it's hard to test. If almost any behaviour can be described as minimising free energy, the theory doesn't tell you much about which behaviour to expect. And it hasn't yet produced a system that does something large and general. **Status in 2026.** An active research community, a strong influence on theories of perception, and early, small-scale AI results. --- ## 7. Evolution and open-endedness: grow it, don't build it **The idea.** Nature produced human intelligence without a designer. It used evolution: variation, selection, and an endless supply of new challenges. So maybe the best way to get AGI isn't to design it, but to build a process that **generates** it. In 2019, Jeff Clune, then at Uber AI Labs, set this out in a paper called **"AI-generating algorithms"**. He argued that hand-designing each piece of intelligence is slow, and proposed three pillars an automatic process would need: **learn the architecture**, **learn the learning algorithm itself**, and **generate the training environments**. **Best result.** That third pillar has the clearest demos. - **POET** (Paired Open-Ended Trailblazer, 2019), by Rui Wang, Joel Lehman, Jeff Clune and Kenneth Stanley, invented its own obstacle courses for a simulated two-legged walker: stumps, gaps, stairs, rough ground. It trained walkers on them, and let walkers move between courses. The result: walkers solved courses that **could not be solved by training on them directly**. Skills learned on easier courses became stepping stones. - The **Darwin Gödel Machine** (2025), from Clune's lab and Sakana AI, applied the same thinking to coding agents. It kept an archive of agents that rewrote their own code and tools. Its score on SWE-bench, a benchmark of real GitHub issues, rose from **20% to 50%** on the authors' own runs. Some "weaker" ancestors turned out to be the source of later breakthroughs, the open-endedness lesson again. **The wall.** Evolution took billions of years and a planet's worth of parallel experiments. Even clever shortcuts are expensive. And open-ended systems are hard to judge: how do you score "interesting"? **Status in 2026.** No longer a fringe idea. Self-generated training tasks and self-improving agents are now standard topics, and the Darwin Gödel Machine was accepted at ICLR 2026. It's increasingly a technique layered *on top of* language models rather than a replacement. --- ## 8. Developmental and embodied robotics: raise it like a child **The idea.** Intelligence isn't something you load into a brain. It grows, through a body, over time, the way a child's does. In 1991, Rodney Brooks at MIT published "Intelligence without representation", arguing that AI had stalled by trying to build elaborate internal models of the world. Build simple behaviours, tightly coupled to sensing and acting, and let more complex ones grow on top. His robots didn't plan much. They just responded to the world, and they worked. Developmental robotics took the next step: build a robot child and let it learn. The EU-funded **RobotCub** project started in 2004 and built the **iCub**, a humanoid the size of a three-and-a-half-year-old, as an open platform for research on embodied cognition. The project produced at least 15 robots for labs around the world. **Best result.** Brooks's ideas went straight into practical robots: he co-founded iRobot, the company behind the Roomba. Developmental robotics produced a large body of research on how robots learn to reach, grasp and imitate. **The wall.** Real-world learning is **slow and expensive**. A robot can't run a million trials overnight like a simulator. Robots break. And no one has raised a robot into anything like a general mind. **Status in 2026.** Embodiment is back in fashion, with a twist: most robotics efforts now attach large pre-trained models to bodies rather than growing a mind from scratch. --- ## 9. AIXI: the perfect agent you can't build **The idea.** In 2000, Marcus Hutter defined **AIXI**, a mathematical description of the best possible agent. It combines two ideas. One is **Solomonoff induction**, a formal recipe for prediction: consider every possible computer program that could explain what you've seen, and trust the shorter ones more. The other is decision theory: pick the action with the highest expected reward. Hutter argued that the result behaves optimally in any computable environment. **Best result.** It's a proof, not a program. AIXI gives researchers a precise answer to "what would perfect general intelligence look like?", and it influenced a generation of AGI and AI safety thinkers. Hutter's student Shane Legg went on to co-found DeepMind. **The wall.** AIXI is **incomputable**. You can't run it, even in principle, because it requires considering infinitely many programs. Later work by Jan Leike and Hutter showed it can't even be approximated as a whole with finite computation, though limited versions can be. **Status in 2026.** A reference point. Some researchers see next-token prediction as a rough, practical cousin of Solomonoff's recipe. That's an interpretation, not a proof. --- ## What the other bets have in common Step back and a pattern shows up. **Almost every alternative bet attacks the same weak spots in today's language models:** - **Learning continuously.** Thousand Brains, active inference and developmental robots all learn as they go. Most deployed language models are frozen after training. - **Learning from little data.** Brains learn a new object in a few touches. LLMs need oceans of text. - **The physical world.** World models, embodied robots and brain emulation all start from bodies and senses, not text. - **Certainty.** Neurosymbolic systems can **prove** an answer is correct, not just make it likely. And almost every one has been **partly absorbed** by the mainstream instead of replacing it. Agents borrowed the cognitive architecture diagram. Maths systems borrowed formal checkers. Coding agents borrowed open-ended search. The likely future is a language-model core with these older ideas attached, roughly what the neurosymbolic camp predicted all along. The last part of this series asks where all of this leaves us in October 2026. --- ## Scoreboard | The bet | What was tried | Best result | Where it hit a wall | Status in 2026 | |---|---|---|---|---| | Cognitive architectures | Soar, ACT-R: one fixed design for the whole mind | TacAir-Soar flew all US fixed-wing aircraft in a 48-hour 1997 simulation; ACT-R-based algebra tutor showed a modest gain in a 17,000-student trial | Knowledge still written by hand for each domain | Maintained, used in psychology; ideas live on in agent design | | Neurosymbolic AI | Language models plus symbolic engines and formal proof checkers | AlphaProof plus AlphaGeometry 2: IMO 2024 silver, 28/42 | Hand translation into formal language; few domains have formal tools | Overtaken at the IMO by pure language models in 2025, absorbed as "write then verify" | | World models and JEPA | Predict abstract summaries of the world, not pixels or tokens | V-JEPA 2 (2025): zero-shot pick-and-place on robot arms in new labs | Not yet broadly general | LeCun's AMI Labs, one of the best-funded alternatives | | Whole brain emulation | Scan a brain, simulate every neuron | Full fruit fly connectome (2024): 139,255 neurons; simple simulation predicted feeding circuits | Human brain is over 600,000 times bigger; wiring is not behaviour | Real neuroscience, very distant AGI path | | Thousand Brains | Copy cortical columns, reference frames and voting | Monty (2025): fast, continuous 3D object learning with little compute | Tiny scale, narrow tests | Open-source research, not production-ready | | Active inference | Minimise surprise through perception and action | Major neuroscience theory; small game results claimed by a company | Hard to test; no large general system | Active niche | | Evolution and open-endedness | Generate environments and agents automatically | POET solved courses direct training couldn't; Darwin Gödel Machine went 20% to 50% on SWE-bench | Enormous compute; hard to score "interesting" | Growing, layered on top of LLMs | | Developmental robotics | Robots that learn through a body, like a child | Brooks's behaviour-based robots; iCub research platform | Real-world learning is slow, costly, fragile | Embodiment back in fashion, via pre-trained models | | AIXI | A mathematical definition of the perfect agent | A precise target for "general intelligence" | Incomputable, cannot be built | A theoretical yardstick | --- ## Key ideas - **The five big bets weren't the only ones.** At least nine other approaches aimed at general intelligence by other roads. - **Neurosymbolic AI had a great 2024**, with an IMO silver from a language model paired with formal proof checking, and then saw pure language models take gold in 2025 without the symbolic half. - **Brain mapping has made real progress**: the full fruit fly wiring diagram in 2024 and a cubic millimetre of mouse cortex in 2025. A human brain is still out of reach. - **Most alternatives target the same LLM weak spots**: continuous learning, learning from little data, the physical world, and certainty. - **The usual fate of an alternative bet is absorption, not victory.** Its best idea gets attached to the mainstream system. - **AIXI is the reminder that "perfect" intelligence is well defined mathematically and impossible to build.** *Sources: [Google DeepMind, AlphaGeometry (January 2024)](https://deepmind.google/discover/blog/alphageometry-an-olympiad-level-ai-system-for-geometry/); [Google DeepMind, AI achieves silver-medal standard at IMO (July 2024)](https://deepmind.google/discover/blog/ai-solves-imo-problems-at-silver-medal-level/); [AlphaGeometry2 paper (arXiv, 2025)](https://arxiv.org/abs/2502.03544); [Gary Marcus on AlphaGeometry2](https://garymarcus.substack.com/p/alphageometry2-impressive-accomplishment); [Google DeepMind, Gemini Deep Think IMO 2025 gold](https://deepmind.google/discover/blog/advanced-version-of-gemini-with-deep-think-officially-achieves-gold-medal-standard-at-the-international-mathematical-olympiad); [Soar (Wikipedia)](https://en.wikipedia.org/wiki/Soar_(cognitive_architecture)); [ACT-R (Wikipedia)](https://en.wikipedia.org/wiki/ACT-R); [Jones, Laird et al., Automated Intelligent Pilots for Combat Flight Simulation (AAAI 1998)](https://mlanthology.org/aaai/1998/jones1998aaai-automated); [Education Week on the RAND Cognitive Tutor trial](https://www.edweek.org/education/study-hybrid-algebra-program-nearly-doubled-math-learning/2013/06); [Sandberg and Bostrom, Whole Brain Emulation: A Roadmap (2008)](https://www.futureofhumanityinstitute.org/s/2008-3.pdf); [Dorkenwald et al., Neuronal wiring diagram of an adult brain (Nature, 2024)](https://pubmed.ncbi.nlm.nih.gov/39358518/); [Shiu et al., A Drosophila computational brain model (Nature, 2024)](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11446845/); [NIH BRAIN Initiative on MICrONS (2025)](https://braininitiative.nih.gov/news-events/blog/brain-director-how-cubic-millimeter-brain-can-change-neuroscience); [Numenta, Thousand Brains Project launch (November 2024)](https://www.businesswire.com/news/home/20241120582696/en/Numenta-Launches-Groundbreaking-Thousand-Brains-Project-Provides-Open-Source-Sensorimotor-Learning-Framework-to-Power-a-Fundamentally-Different-Approach-To-AI); [Thousand-Brains Systems paper (arXiv, 2025)](https://arxiv.org/abs/2507.04494); [Monty on GitHub](https://github.com/thousandbrainsproject/tbp.monty); [Friston, The free-energy principle: a unified brain theory? (2010)](https://www.nature.com/articles/nrn2787); [VERSES AI, AXIOM press release (June 2025)](https://www.globenewswire.com/news-release/2025/06/02/3091981/0/en/VERSES-Digital-Brain-Beats-Google-s-Top-AI-At-Gameworld-10k-Atari-Challenge.html); [Clune, AI-GAs (arXiv, 2019)](https://arxiv.org/abs/1905.10985); [Wang et al., POET (arXiv, 2019)](https://arxiv.org/abs/1901.01753); [Sakana AI, Darwin Gödel Machine](https://sakana.ai/dgm/); [Brooks, Intelligence without representation (1991)](https://cse.buffalo.edu/~rapaport/663/brooks.html); [iCub history](https://icub.iit.it/about-us/icub-history); [Hutter, Towards a Universal Theory of Artificial Intelligence (2000)](https://arxiv.org/abs/cs/0012011); [Leike and Hutter, On the computability of Solomonoff induction and AIXI](https://openresearch-repository.anu.edu.au/items/28832ba1-2dc1-4a98-b164-b0fc7c00ad01).* *Related: [Bet #5: Teach It to Think (Reasoning and Agents)](https://phamphutai.com/2026/10/21/reasoning-and-agents-teach-it-to-think/), [AGI: Where We Are Now (October 2026)](https://phamphutai.com/2026/10/21/agi-where-we-are-now/), and [World Models: Can AI Simulate Reality?](https://phamphutai.com/2026/10/20/world-models/).*