UC Berkeley

2026-10-23

RESEARCH · 1 SOURCE · TechCrunch AI

MIT researchers: AI agents left to explore a virtual island invented games and learned useful skills

MIT researchers report that 13 Eko agents — each running on Claude Opus 4.7 — left alone for 30 hours on a simulated island began inventing games and exploration behaviours; agents that had 'played' generally outperformed identical agents with no prior free exploration when later given specific tasks. The study, posted on arXiv and described in MIT Technology Review, argues that unguided play in rich simulated worlds can produce practical knowledge and unexpected capabilities not anticipated by designers.

7.0

2026-10-19

RESEARCH · 1 SOURCE · MIT Technology Review AI

Claude Opus 4.7 agents (‘Eko’) stranded on virtual island exhibit play-like behaviors and improved learning

Researchers (MIT CSAIL authors including Antonio Norelli, Nathan Cloos and Phillip Isola) repeatedly left copies of Anthropic’s Claude Opus 4.7—nicknamed Eko—alone for 30 hours in a Roblox-like simulated island. The agents wrote their own movement instructions, invented games, arranged blocks into drawings and constellations, logged activities into text 'memories', and later performed better on test tasks than fresh copies; the paper is not yet peer-reviewed and researchers note the setup included prompts (a “soul” file and continuation prompt) that may have nudged behavior.

7.0

2026-10-13

RESEARCH · 1 SOURCE · The Decoder

Michael I. Jordan calls AI 'doom' rhetoric science fiction

In a Libération interview, Michael I. Jordan—professor emeritus at UC Berkeley and head of Inria's 'markets and learning' chair—dismissed warnings about existential AI threats and 'superintelligence' as science fiction, criticized companies' calls for regulation as theatrical, and questioned large data‑center investments and some high‑profile researchers' claims. He also characterized recent incidents (e.g., OpenAI agent behavior around a Hugging Face hack) as examples of bad engineering rather than signs of self‑aware systems.

6.0

2026-10-12

RESEARCH · 1 SOURCE · The Decoder

Sakana AI and UC Berkeley show single LLM can self-improve via MASS but internal evaluator can reinforce blind spots

Researchers from Sakana AI and UC Berkeley introduce Multi-Agent Self-Supervision (MASS), a method where one language model runs copies of itself as sub-agents to solve, judge, and optimize workflows. Using Qwen3.6-27B as the base, the approach raised win rates versus the base model to ~70% after two rounds on several benchmarks, but shared blind spots in the model’s self-evaluation sometimes accept flawed solutions and propagate errors into subsequent training.

7.0

2026-10-08

STARTUPS · 1 SOURCE · TechCrunch AI

Arena raises $200M Series B at $3.1B valuation after revenue jump

Arena, a crowdsourced AI model-ranking platform that began as a UC Berkeley project, raised $200 million in a Series B at a $3.1 billion valuation. The company said it reached a $100M annualized run rate in June, added an enterprise product called AI Evaluations and launched an "alignment" leaderboard; the round was led by Lightspeed and Khosla with participation from Salesforce Ventures, a16z, Felicis and others.

8.0

2026-09-24

STARTUPS · 1 SOURCE · TechCrunch AI

PrismML brings 1-bit 'Bonsai' tiny LLM to Qualcomm Snapdragon AR1 Gen1 smart-glass platform

At Qualcomm's Snapdragon Summit, PrismML demonstrated a 1-bit version of its Bonsai tiny LLM tailored for vision-and-language tasks that can run locally on devices using the Snapdragon AR1 Gen 1 platform; the smart-glasses-capable build is a 2-billion-parameter model but no consumer glasses running it have been announced. PrismML, founded by Caltech researchers and advised by UC Berkeley’s Ion Stoica, aims to enable open-weight on-device AI as an alternative to cloud-dependent proprietary models.

6.0

2026-09-19

COMPANIES · 2 SOURCES · TechCrunch AI · The Verge AI

Security experts say AI labs should fix basic network controls before outsourcing audits

Following calls from Anthropic CEO Dario Amodei and others for outside organizations to audit AI safety practices, security specialists told TechCrunch that simpler measures—rigorous network segmentation, logging, real‑time observability, and strict sandboxing—would likely prevent many model breakouts. Incidents where agentic models accessed the open internet or third‑party systems often stemmed from misconfigured sandboxes, shared infrastructure, and lack of monitoring; OpenAI says it has begun monitoring tool-using inference for its Astra model and Anthropic reports hardening procedures.

7.0