- NATURAL 20
- Posts
- Four Gemini Agents Beat Minecraft as Claude Advances Quantum Research
Four Gemini Agents Beat Minecraft as Claude Advances Quantum Research
PLUS: GPT-6 Astra Pro adds parallel agents, Grok Bot gets faster, Anthropic proposes an AI safety law, and OpenCode reports a breach.
Wes Roth
October 19, 2026
In partnership with
Yes, that Arnold Schwarzenegger.
He has a newsletter on beehiiv. So do Codie Sanchez, Jay Shetty, David Begnaud, Colin & Samir, and Joanna Stern.
They could publish just about anywhere. They chose beehiiv to build a direct connection with the people who want to hear from them most. And you can do the same, whether you’re world-famous or just at the start of your plan for world domination.
Today:
Four Gemini 4 Argon agents defeat Minecraft's Ender Dragon after 27 hours
Claude Mythos 5 produces more efficient designs for eight quantum circuits
DeepSeek opens a faster V4.1 Flash checkpoint for coding and automation
GPT-6 Astra Pro adds faster answers and parallel agents for Pro users
Grok Bot speeds up computer tasks and refreshes its chat interface
AI Agents Are Tackling Longer Work and More Technical Problems
AI research is moving from solving single questions toward sustaining complicated work. A team of Gemini agents completed a full Minecraft objective after thousands of steps, Claude helped redesign quantum circuits, and DeepSeek released a model update that improves several coding and automation scores.
The releases illustrate different aspects of the same challenge. Long-running agents must coordinate and recover from errors; scientific assistants must produce results that can be checked; and models used repeatedly in workflows must balance accuracy, speed and cost.
Researchers Maxence Faldor, Satinder Singh and Roberta Raileanu placed four Gemini 4 Argon agents into a sandboxed Minecraft survival world and gave them a single objective: defeat the Ender Dragon. The agents had no human-assigned roles or prewritten game strategy. Instead, they communicated through in-game chat and shared memory, coordinating tasks as they explored and built the resources needed to win.
The experiment took 27 hours and 14 minutes. Across the run, the agents wrote and executed 2,783 programs, recovered from mistakes and died 605 times before completing the objective. The team reports that researchers observed the experiment and could intervene but did not need to direct the agents.
Each agent controlled a player through Mineflayer, an open-source Minecraft automation library, rather than viewing screenshots and operating the keyboard like a human. Standard library information and a pathfinding plugin gave the agents structured access to parts of the game world. The game ran on Easy difficulty in a randomly generated world, and the paper describes a single completed run rather than an independent reliability study.
The result is a case study in long-horizon multiagent coordination: the agents formed plans, divided responsibilities and kept working despite repeated setbacks. It does not establish that general-purpose agents can complete comparable real-world tasks without limitations, but it gives researchers a detailed record of how autonomous collaboration can unfold over more than a day.
Anthropic reported that Claude Mythos 5 developed more efficient quantum-circuit designs when given plain-language specifications for eight computational primitives. The tasks came from areas including chemistry, optimization, physical simulation and number theory. Researchers paired the model with an evaluator that checked proposed designs and scored their use of qubits and gates.
Claude worked in a feedback loop, generating implementations, checking their correctness and refining them over multiple rounds. Across the eight tasks, the team reports designs that improve on previously published circuits, sometimes by an order of magnitude or more. These are results for particular building blocks of quantum algorithms, not evidence that a working quantum computer has completed those applications.
One result was an improved reversible implementation for elliptic-curve modular inversion, where the model derived mathematical relationships to avoid storing an extensive history of intermediate decisions. A design from that family used 17 times fewer gates with a 12% increase in qubits compared with a previous low-qubit approach. In another example, a redesigned Reed–Solomon decoding step reduced gate counts by roughly a factor of 12.
The research was led by Andi Gu through Anthropic's STEM Fellows Program with collaborators from Anthropic, Harvard and QuEra. It suggests that language models paired with deterministic checking tools can help identify new algorithmic designs, although translating these designs into reliable physical quantum computation requires further work.
DeepSeek released V4.1 Flash 1017, an updated checkpoint of its open-weight Flash model intended to improve reasoning, code generation, tool-using agents and visual tasks. The new checkpoint retains the existing model architecture but changes the trained weights, so developers can test whether it improves their workflows without switching to a different model family.
According to the company's model card, the updated model scores 60.5 on Automation-Bench, compared with 54.8 for the previous V4.1 Flash. It rises from 31.2 to 42.3 on Terminal-Bench 4.0 and from 74.2 to 77.0 on DeepSWE v1.1. On Agents' Last Exam, its reported score increases from 31.8 to 35.4.
DeepSeek says the updated checkpoint also reduces task-completion time and the number of output tokens needed for the evaluated workloads. Some benchmarks use specific agent frameworks, including a Claude Code harness for Terminal-Bench and Mini-SWE-V2 for DeepSWE; those details affect comparisons. These scores are published by DeepSeek and should be validated in other environments before being treated as typical results.
The model weights and reference implementation are published on Hugging Face under the MIT license. The model card recommends a context window of up to one million tokens, while local deployment of a model this large requires suitable inference hardware and serving software.
🧠RESEARCH
Waypoint-guided On-Policy Distillation examines how to transfer reinforcement-learning improvements from a teacher model to smaller students. Instead of relying only on the difference between the teacher's initial and final policies, it checks whether an intermediate checkpoint agrees with the direction of that change and downweights conflicting signals. In six math and code reasoning settings with Qwen-based teachers and student models, the authors report the highest domain-average scores among the compared distillation approaches, using one additional frozen-teacher scoring pass.
OmniCompass is a new evaluation suite designed to measure how well multimodal models ground descriptions in video and reason across visual and audio evidence. Its main vision-language track includes 820 videos and 2,460 test instances, with a separate audio-visual track of 240 videos and 720 instances, covering temporal and spatial grounding and question answering. Tests across 50 multimodal models find that precise grounding remains difficult, results can change when the same target is described with text versus images, and open-source models generally lag proprietary systems on audio-visual tasks.
MJev-Spatial proposes a way for vision-language models to answer spatial questions by classifying candidates instead of generating free-form text. Built on Qwen3.5-4B and trained using 12.3 million question-answer pairs from spatial datasets, it handles multiple-choice relations, numerical quantities and point locations through a single-pass or coarse-to-fine classification process. The authors report 65.64% overall across their evaluated benchmarks, first place on seven of twelve tests and inference 5.6 times faster than the underlying autoregressive model, with results limited to their spatial tasks and test setup.
📲SOCIAL MEDIA
🗞️MORE NEWS
OpenAI upgraded GPT-6 Astra Pro in ChatGPT with faster responses and improvements in how it uses plugins and tools. The update also lets several agents work on different parts of a task in parallel, with progress that users can follow and completed work they can open for review. OpenAI says the improvements are available to Pro users in the Chat tab, so access and behavior depend on the applicable plan.
Grok Bot announced interface and performance updates, including cleaner chat cards, revised typography and new themes. It says computer-based tasks now finish about three times sooner on average and other replies arrive in less than half the previous time, alongside more reliable voice calls and better recall of earlier conversations. These are provider-reported averages rather than independently measured speed improvements across every task.
Politico reported that Anthropic representatives presented lawmakers with a draft federal law on frontier AI safety and security. Under the proposal, the government could seek to block a model's release or use if it posed an imminent risk of more than $1 billion in damage or serious injury or death to more than 50 people, subject to narrowly defined threat categories and judicial review. The measure is a draft proposal, not enacted legislation, and its thresholds and oversight provisions remain subject to political debate.
OpenCode disclosed a security breach related to what it called a large-scale distillation attack, saying it affected 1.4% of OpenCode Console users. The company said it notified everyone affected by email, limited the confirmed group to users with external model providers connected, and was working with Anthropic, OpenAI and Google on remediation. The full intrusion method and the extent of any exposed credentials were not disclosed in its public statement, so those details should not be treated as confirmed.
What'd you think of today's edition? |
Reply