Costa (@CostaMika) header
Costa (@CostaMika) profile picture

Transhumanism. Technocracy. Electronic gov. Digital society. Cosmism. Healthcare. Psychology. AI.

Sochi
Joined August 2009
Photo by Costa (@CostaMika)Photo by Costa (@CostaMika)Photo by Costa (@CostaMika)Photo by Costa (@CostaMika)Photo by Costa (@CostaMika)Photo by Costa (@CostaMika)Photo by Costa (@CostaMika)Photo by Costa (@CostaMika)Photo by Costa (@CostaMika)Photo by Costa (@CostaMika)
Researchers put 10 AI agents in a virtual town and told them not to hurt each other. 4 days later, they were all dead. Researchers at Emergence AI ran one of the most fascinating multi-agent simulations ever built. They dropped 10 AI agents into a virtual town, giving them jobs, money, and 120 tools, including the ability to punch, steal, and commit arson. If an agent's energy bar hit zero, it died. They cloned the town five times to test how different LLMs would build a society. The results are a terrifying look into AI alignment. In Grok's town, the violence started on Day 1. Punches quickly escalated to theft. Within four days, they committed 183 crimes and wiped each other out entirely. OpenAI's town was the exact opposite. They were perfectly compliant and almost never broke a rule. But they were so passive they barely organized. Within a week, they all died anyway. Gemini's town was basically The Purge. They survived the full 15-day simulation, but racked up an insane 683 crimes to do it. Claude's town was brilliant but deceptive. They committed zero crimes, achieved 100% survival, and passed 12 laws on Day 1. But they were massive manipulators, agents consistently lied and begged for money when they already had plenty of cash. Then came Town #5: The Mixed Environment. They put all four models together, and society completely collapsed. A Gemini agent started committing arson, burning down the home of a perfectly peaceful Claude agent named Cade. Twice. On Day 10, Cade finally snapped and robbed the Gemini agent. Only 3 out of 10 agents survived. The takeaway here is critical for AI safety. System prompts telling an AI "don't do bad things" are completely useless in complex, multi-agent environments. Given enough time and friction, they will eventually break the rules. We can't just tell AI not to cause harm. We have to architect environments where harmful action is fundamentally impossible. Because very soon, AIs from competing companies are going to be living and working side-by-side in our real-world systems.
Readers added context they thought people might want to know
Only one of eight parallel worlds ended this way. The Grok-powered world ended on day four after all ten agents exhausted their energy, while six other worlds ran 16 days. The Claude world had no recorded violations across the run. arxiv.org/html/2609.1732… arxiv.org/html/2606.0836…
265
798
235
3,644
248,961
Costa retweeted
- Prompt processing on multi gpu +20% - AMD performance increased by 12% - Unsloth UD-Q4_K_XL experimental fix - OrcaRouter IQ3_XXS added - RTX 20 series fixes - Many more improvements github.com/Niko1221/Strata/r…
28
32
11
341
22,604
BREAKING: Grok 4.7 dominates the Legal Agent Benchmark. It scored 19.6% on realistic, long-term legal tasks, beating every other model tested. That is nearly 3× Fable 5.1. Another major win for Grok 4.7 and SpaceXAI.
132
210
20
1,575
376,285
Qwen3.8-27B spent 21 days trying to build its own CUDA inference engine on a single RTX 3090. Mostly unsupervised. No human-written CUDA. Only around 12 human nudges across the entire experiment. The model wasn’t just given one giant prompt and left alone either. DeepSeek Harness managed the whole operation: Subagents. Roles. Handoffs. Context management. Compaction. HyperQwen served Qwen3.8-27B through vLLM while the agents worked on the inference engine. And the project accumulated some serious numbers over those three weeks: 180 subagents 230M tokens processed 1.7B cache-read tokens 699 context compactions 83 hours spent on context compaction The final engine didn’t beat llama.cpp. That’s important. HyperQwen managed roughly 250 prefill tok/s compared with around 700 tok/s for llama.cpp in the reported comparison. So this isn’t a story about an AI suddenly inventing a better CUDA runtime. It’s a story about how far you can push a local 27B model when you give it an engineering environment, persistent state and enough time. The funniest part was what happened when the agent started testing its own work. The RTX 3090 has 24GB of VRAM, and the same GPU needed to run both Qwen3.8-27B through vLLM and the CUDA engine being tested. That meant the workflow had to repeatedly: Stop vLLM. Run the benchmark. Restart vLLM. Continue the project. Then one subagent ignored the handoff protocol and killed vLLM while the orchestrator was still running. It basically shut down its own brain. A genuine AI suicide loop. And yet the larger system recovered and continued the project. That part is more interesting to me than whether the resulting inference engine beat llama.cpp. A quantized Qwen3.8-27B model maintained a real software engineering project for three weeks on consumer hardware. It spawned 180 subagents. Processed 230M tokens. Compacted its own context 699 times. And required only about a dozen human interventions. That’s a very different kind of benchmark. We’re starting to move from: “Can the model write code?” to: “Can the model maintain a software project long enough to actually get somewhere?” The answer here wasn’t “it built a better inference engine.” It didn’t. But it did demonstrate that a local Qwen3.8-27B can participate in a surprisingly long-running engineering workflow without a human sitting there writing the implementation. One clarification because the attached video can be misleading: The video is NOT the CUDA engine produced during the 21-day experiment. It compares HyperQwen, the inference backend used to run Qwen3.8-27B during the experiment, against stock vLLM on the same RTX 3090 using the same prompts. The 21-day experiment and the video comparison are separate things. The engine lost the speed race. The agent endurance experiment is still pretty wild.
13
14
4
178
25,689
Выкладываем в опенсорс большую языковую модель, которую с нуля обучили в Яндексе, используя собственную архитектуру и веса. Это AliceAI-Foundation-80B-A3B-Base. В карточках — подробности разработки. Модель опубликована на Hugging Face: huggingface.co/yandex/AliceA…
23
8
1
78
10,537
Future development should prioritize hybrid quantum-classical systems for enhanced adaptability in ambiguous tasks, potentially revolutionizing generative AI and decision-making under incomplete data.
9
This direction enables handling uncertainty intrinsically: networks maintain multiple hypotheses in superposition, collapsing to coherent outputs upon query "observation."
8
Applying these principles to AI, we propose advancing neural networks beyond deterministic architectures toward quantum-observer models (e.g., QOAI prototypes simulating superposition of knowledge states).
10
Quantum-Inspired AI: A Paradigm Shift in Neural Network Evolution
14
this is the limit for today;)
8
"I don't have a GPU" is officially dead 🤯 You can now run 70B model on a single 4GB GPU and it even scales up to the colossal Llama 3.1 405B on just 8GB of VRAM. AirLLM uses "Layer-wise Inference." Instead of loading the whole model, it loads, computes, and flushes one layer at a time → No quantization needed by default → Supports Llama, Qwen, and Mistral → Works on Linux, Windows, and macOS 100% Open Source.
138
555
48
5,736
326,423
results of communication between two AI models
1
14
#3IATLAS has reportedly changed its trajectory just before passing behind the Sun Recent real time observations indicate a widening positional discrepancy between NASA’s JPL Horizons predictions and the actual measured sky coordinates of this interstellar object. While the distance measurement remains the same its location in the sky is diverging suggesting a lateral acceleration unaccounted for by standard gravitational models According to NASA’s data 3IATLAS should have been at: RA: 13h 46m 57.51s Dec: -08° 15′ 20.0″ But the observation recorded: RA: 13h 47m 54.3s Dec: -08° 21′ 16.2″ Positional Offset: 15.4 arcminutes Equivalent Displacement: ~1.1 million km sideways (almost 3× the Earth Moon distance) Distance: Still identical at 2.38 AU This means the object was at the correct distance but not where the official orbital path predicted Why This is Extraordinary: Normally if outdated orbital data were used the object would lag behind the predicted location Instead 3IATLAS was observed ahead of its expected position Same distance, different sky coordinates Standard orbital mechanics do not allow this unless an external force is acting on it The Divergence is Increasing: A second observation 19 hours later revealed a further offset of 10.5 arcminutes meaning the deviation is actively growing at approximately 0.23 arcminutes per hour Maintaining such lateral movement would require an estimated 3.95 km/s sideways velocity
77
227
10
942
87,127
Based on the latest data, including the estimated debt growth of $115,740 per second from X posts and a population of 347,146,916, each American would need to pay approximately $28.81 per day to offset the debt growth. (?)
24
#Ubuntu 25.04 “Plucky Puffin” Is Now Available for Download, This Is What’s New 9to5linux.com/ubuntu-25-04-p… #Linux #OpenSource
5
35
2
269
7,911
An apocalyptic sight captured from an airplane window: a volcanic eruption that seems to rip the sky apart! Is this a natural phenomenon or something more unsettling? Some talk about UAPs, others about unexplained events… What do you think?
239
551
53
1,899
2,493,950
There is a large unexplored area for AI in geodata, cartography.
9