SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Four Gemini 4 Argon Agents Complete Minecraft's Ender Dragon Without Human Guidance

0
·1 views

Google DeepMind placed four Gemini 4 Argon agents in a random Minecraft survival world with no assigned roles or custom strategies, relying only on the standard Mineflayer library. After 27 hours and 14 minutes, they defeated the Ender Dragon, writing 2,783 JavaScript programs and exchanging 229 chat messages along the way. The agents died 605 times and spontaneously divided responsibilities, with one taking on cleanup work, while a fifth observer agent reviewed the run. The session consumed about 1.35 billion input tokens, costing roughly $1,620 at standard pricing. The researchers noted caveats including a single run on easy difficulty, the model's likely prior exposure to Minecraft in training data, and an unknown repeatability rate.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Seeded inbox test clears spam placement; signup flow overhauled with free credits

A team that sent 106 cold emails with zero replies ran a seeded deliverability test after community feedback, sending the same message through two paths. Both copies reached the inbox within two minutes, with no spam-folder placements observed. The team notes that a seed test cannot show how every outside filter would score the message. Having ruled out deliverability, they turned to the signup page, which had led with a form and offered nothing free. The revised flow adds one-click Google signup, gives new accounts 1,000 credits with no card required, and keeps those credits valid for 365 days. The team plans to measure whether these changes affect reply rates on a new, separately counted outreach batch.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Stanford's Hazy Research Proposes Next-Concept Modeling as Alternative to Token-by-Token AI Generation

Researchers at Stanford University's Hazy Research group have introduced Next-Concept Modeling, a new language modeling architecture released in October 2026. Unlike traditional autoregressive LLMs, which generate text one token at a time, the approach first builds a global representation of the response and then refines it in parallel. It uses a multiscale hierarchy of discrete codes, predicting all codes at each level simultaneously and conditioning each subsequent level on prior outputs. The reported results include 10 to 31 times faster inference than evaluated diffusion baselines and roughly 4 to 5 times fewer training tokens than certain baselines. The researchers note these results do not establish performance advantages over widely used models such as GPT, Qwen, or Llama.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

One Spot-Checked Listing Undermined an AI-Built List of 261 'No Website' Businesses

A developer offering one-page websites to small Turkish businesses used an AI agent to compile 261 candidates with good ratings and no real website. A random verification check, run in a separate session, found that a vet clinic marked 'no website' actually had a WhatsApp link in its Google Maps website field. The developer concluded that the status column likely reflected only the data the agent read, not what a visitor sees. Since a single sample cannot measure the error rate, the developer decided the whole column could not be trusted without further verification. The post describes a standing rule of spot-checking one random claim per delivery before accepting the rest.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Unintended Details in a Locked Reference Image Carry Into Every Output

A recent test found that a reference image preserves everything visible in it, not just the traits a creator deliberately specified. A character generated from a brief with few details unexpectedly included a small hair clip that was never requested and appeared in only one variant. That image was chosen as the master reference for later work, and the stray clip persisted in a downstream layout tool whose prompt never mentioned it. The author argues creators should check master images for incidental details and fix unwanted ones before locking the reference, and should document any accidents worth keeping in written handoff notes. The author also notes the test is a single observation, not a guarantee, since small details can drift.

0 comments Read more at DEV Community