“Brain cells beat AI”: Who the biocomputers actually beat
Since the weekend, the story has been making the rounds again: 15 networked “mini-brains” from Cortical Labs supposedly beat an AI, 70 to 32 percent. We read the underlying paper. The opponent was a learning algorithm with eight artificial neurons, no experience memory and just 150 moves. Given more time, it reaches around 75 percent on its own.

What this is about
“Quince minicerebros superan a una IA” (“Fifteen mini-brains beat an AI”), Spanish news sites ran on November 21. Shortly afterward, German-language sites picked up the story again. Australian company Cortical Labs has linked 15 of its CL1 biocomputers into a cluster called HIVE. A CL1 is a chip on which living human neurons grow, stimulated and recorded through electrodes. In a navigation task, HIVE reportedly reached about 70 percent of the maximum score, while a “conventional AI” managed only 32 percent.
We wanted to know what this win rests on. So we read the underlying research paper and the earlier study on the game Pong, and compared the power figures. The short version: the cells won because the contest was very short and the opponent very small. Still, the cells do show something interesting.
What HIVE claims and what's backed up
Cortical Labs presented the HIVE results on October 26 and 27, 2026, at the first International Conference on Organoid Intelligence in Washington. New Atlas reports these numbers: in a symbol recognition task, seven linked devices raised accuracy from about 80 to 99 percent. In grid navigation, a Deep Q Network (DQN) reached about 32 percent and the best single culture 40 percent. Multiple cultures voting on every move reached about 70 percent, and groups specialized in subtasks about 80 percent. The source is Cortical Labs itself. New Atlas explicitly calls the results “preliminary.” We couldn't find a paper, slides or an abstract for the talk. On the conference website, Cortical Labs is listed only as a sponsor.
The foundation, however, is documented: the preprint “Embodied Neurocomputation” (arXiv 2605.13315). Version 1 came out on May 13, 2026, version 2 on October 2. Almost all authors work at Cortical Labs; Bernhard Sendhoff is from the Honda Research Institute Europe. In it, neuron cultures steer a figure through a 6×6 grid world with walls toward “food.” The numbers line up: 40 divided by 32 is exactly 1.25, and that's the factor the preprint gives for the cells' lead in the condition with five short episodes. Our guess: the DQN figure in the HIVE talk comes from the same setup. That hasn't been confirmed.
| Condition | Source | Result |
|---|---|---|
| 1 episode of 150 steps | Preprint | Cells 1.18× DQN, 3.6× no-learning control |
| 5 episodes of 30 steps, 2 min. break | Preprint | Cells 1.25× DQN, 5.6× control |
| DQN with up to 1,000 episodes | Preprint | DQN ≈ 0.75 of maximum score |
| HIVE, short budget | Talk | DQN 32 %, single culture 40 %, voting 70 %, specialized groups ≈ 80 % |
The opponent: a mini DQN with eight neurons
The Deep Q Network goes back to a 2013 DeepMind paper (arXiv 1312.5602). It learns by trial and error how much each action pays off in the long run. It's known for needing lots of attempts. Appendix A.7 of the preprint describes it: a single hidden layer with 8 neurons. The HTML version says “88,” but numbers there are printed twice throughout, and another passage explicitly says “8.” On top of that, the network learns online and uses each experience only once, just like the cell culture.
That strips the DQN of its most important component. Normally it stores experiences in a buffer and learns from them again and again (experience replay). For equal treatment, that may be consistent. For the headline “beats AI,” it's decisive. To be fair: according to the paper, the authors tried around 9,500 settings for learning rate, target network update interval and exploration, tuned to this short time frame. The text doesn't say how many random seeds the DQN figure covers or whether it's a mean or a best value.
The most important sentence is in the appendix too: all DQN variants tested needed “at least four times more episodes” to match the cells' score after five episodes. That's around 600 steps instead of 150. Given enough time, the DQN levels off at about 0.75. That's above HIVE's voting score of 70 percent and just below the specialized groups' 80 percent.
The pattern isn't new. In the earlier Pong study (arXiv 2405.16946, published in 2025 in Cyborg and Bionic Systems), the cultures managed about 70 rallies in 20 minutes. So DQN, A2C and PPO got exactly 70 episodes. According to the publication, the learning algorithms overtook the cells when allowed to train for “tens of thousands” of episodes. Several authors were employed by or contracted to Cortical Labs, and Brett Kagan holds shares.
The task: three sensor values, three actions
The figure gets a single sensor value. −1 means the scent is strongest to the left. 0 means it's strongest straight ahead. +1 means it's strongest to the right or behind. The figure can move forward, turn left or turn right. Food earns +2 points, bumping into a wall −0.2, plus a small bonus when the scent gets stronger. The food appears in the same spot every episode because the random number generator starts the same way each time.
We originally planned to rebuild this world and run comparison methods on the same budget: tabular Q-learning, a fixed rule and random play. We did not run this simulation for this article. So we're not reporting any measurements from it. But a few conclusions follow from the task description alone, no computer required:
- There are only 3 × 3 = 9 combinations of sensor value and action. If you only react to the current sensor value, you have exactly 27 possible fixed strategies to choose from. Tabular Q-learning, which simply keeps a table of experience values, only has to learn nine numbers.
- The obvious rule “left → turn left, ahead → go straight, otherwise → turn right” is exactly what the scent sensor suggests. It doesn't need a single learning step.
- Using a neural network for this table and training it with 150 moves and no experience memory makes the task artificially hard for the silicon.
Whether the fixed rule is optimal, say at walls, would take a replication to settle. Our take: the “beaten” opponent works as a benchmark for the question “How fast does something learn with no prior knowledge at all?” It doesn't work as a stand-in for “AI.”
1,296 attempts, 12 winners
In the first step, the researchers tested 1,296 combinations of six stimulation parameters, including frequency, current and pulse width. A setting counted as a “top performer” if it exceeded the 99th percentile of a simulated random control. Do the math and you'll see a problem: even if not a single setting were any good, a bit over 1 percent of 1,296, or about 13, would clear that threshold by chance alone.
That's why the authors added a second stage. 64 candidates ran four times each, and 12 of them showed “consistent” improvements over the controls. That's a real safeguard. The success rate was still under 1 percent (12 out of 1,296). For the DQN comparison, the best cell configurations then competed, tested on five cultures. So there was cherry-picking on both sides: around 1,300 variants for the cells, around 9,500 for the DQN. What tips the scales isn't the selection, it's the tiny budget.
What the cells do show
If you only criticize the headline, you miss the controls. They're the strongest part of the paper. When the experiment runs with nutrient medium and no cells, the learning curve stays flat (slope +0.0046 per episode). When the readout regions are spatially shuffled, it even drops (−0.0067). With the best settings, on the other hand, agreement with the target rises from 0.39 to 0.53. Chance level here is 0.33. The difference from the worst settings is significant (p < 0.001). Learning kicks in around the third episode. Five short episodes with breaks work much better than one long one (p < 10⁻⁴). Breaking down the results by target variable also shows that a maximum stimulation frequency of 40 to 60 Hz matters most.
For neuroscience, that's valuable: living networks adapt their behavior to feedback within minutes. It's just that each individual move is slow. With the best settings, a step takes 2 to 4 seconds, so 150 steps take 5 to 10 minutes.
The criticism from the research community stands. Back in 2023, around 30 researchers pushed back in Neuron against the “sentience” interpretation of the Pong study. Tony Zador (Cold Spring Harbor) told STAT News that trying to catch up with silicon AI using cell cultures is a dead end. Sergiu Pasca (Stanford) warns in the same piece:
“Overblown claims can confuse the public and policymakers about what these systems can actually do.”
A review in Frontiers from March 2026 speaks of “neurohype.” Long-term learning, it says, hasn't yet been demonstrated in cultures like these.
Power: 25 watts per unit, not 1,000
German-language reports claim a CL1 draws 650 to 1,000 watts. That's apparently a mix-up with the whole rack. The Next Web corrected the same error on August 21, 2026: according to Cortical Labs, one unit draws about 25 watts, and a rack of 20 units 800 to 1,000 watts. The paper measured 67 CL1 units in cloud operation and arrives at a median of 22.7 watts (99th percentile: 25.5 W). That covers the hardware that keeps the culture alive and connected.
Using these figures, we estimated what a 150-step run costs:
| System | Power | Time for 150 steps | Energy (rough estimate) |
|---|---|---|---|
| 1 CL1 | ≈ 22.7 W | 5–10 min. | ≈ 2–4 Wh |
| HIVE (15 CL1) | ≈ 340 W (devices only) | 5–10 min. | ≈ 28–57 Wh |
| DQN with 8 neurons | Laptop CPU | Fraction of a second (estimate) | well under 1 Wh |
The paper compares the biocomputers with GPU workstations drawing 579 to 641 watts. But a network with eight neurons doesn't need a GPU. Our take: for this task, the energy balance favors silicon, not cells. Then there's the cost: $35,000 per unit, $300 per week in the cloud (IEEE Spectrum). And the cells have to be fed every three days and last about six months at most.
The bigger picture
The claim “brain cells beat AI” isn't false, but it's misleading. The cells won against a deliberately small, memoryless learning algorithm from 2013. The contest ran over a distance of 150 moves, in a task with three sensor values. With four to 200 times more episodes, that algorithm catches up or pulls ahead. Some Spanish outlets like Qué! do mention the weak opponent, but misread the 1.18 factor as “in 18 to 25 percent of cases.”
What's real and interesting is something else. Living cultures demonstrably learn from just a little feedback, and the no-cell controls argue against a pure measurement artifact. The networking advantage in HIVE is so far known only from a talk. Our verdict: anyone who wants to seriously compare biocomputers with AI will need three things going forward: the seven-line rule and tabular Q-learning as baselines, a DQN with experience replay, and a real energy measurement per solved task. Until then, “beats AI” is a statement about experimental design, not about intelligence.
Sources
- arXiv: Embodied Neurocomputation (2605.13315) arxiv.org
- arXiv HTML v2: Embodied Neurocomputation arxiv.org
- arXiv HTML v1: Embodied Neurocomputation arxiv.org
- New Atlas: Living computers forge a 'hive mind' newatlas.com
- arXiv: Biological Neurons Compete with Deep RL (Pong) arxiv.org
- PMC: Cyborg and Bionic Systems – Pong-Vergleich pmc.ncbi.nlm.nih.gov
- arXiv: Playing Atari with Deep Reinforcement Learning (DQN, 2013) arxiv.org
- STAT News: Brain organoid pioneers fear backlash statnews.com
- Frontiers in Molecular Neuroscience: Integrating neural organoids and AI frontiersin.org
- IEEE Spectrum: Biological computer for sale spectrum.ieee.org
- The Next Web: Singapore biological data centre (mit Korrektur) thenextweb.com
- Qué!: Cortical Labs conecta 15 minicerebros que.es
- El Periódico Digital: Quince minicerebros se conectan y superan a una IA elperiodico.digital
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.