The Frontier - SiliconNoon (Page 12)

The Frontier

Bonsai 2 27B puts a 27B model on a laptop in 5.9 GB

Two releases worth your attention today, and they pull in opposite directions: one makes a 27B model fit where it never fit before, the other shows the numbers we use to rank models are softer than advertised. PrismML shipped Ternary Bonsai 2 27B, a compressed build of Qwen3.8-27B that holds more than 98 per cent of the full-precision model's capability at a fraction of its size. The language weights land at 5.9 GB — the on-disk pack is 8.60 GB once the 0.92 GB vision tower is counted — which i

The Frontier

Union Alpha burned 2 billion tokens on day one — and nobody claims it

OpenRouter's newest stealth model has no name, no model card, and no published benchmark table — only a free week, a 262K context window, and a lot of curious developers. A model nobody will put a name to burned roughly 2 billion tokens in its first day on OpenRouter, and more than 100 billion in total since, according to the platform's own telemetry. Union Alpha landed on the router on September 16 as a stealth listing: text and image input, a 262,144-token context window, up to 131,072 tokens

The Frontier

An AI has won the Metaculus Cup for the first time

Two stories this hour about machines doing human work in public, and one of them came with a scoreboard. An AI system won the seasonal Metaculus Cup on September 5 — the first time a bot has taken the top spot in the forecasting tournament that humans have owned since it began. The Economist reports that AI entries also took second and fifth place, pushing humans into third and fourth. Hundreds of entrants had to call questions resolving by September: whether an American state or EU country wou

The Frontier

Astra played Minecraft for 141 hours — then a creeper erased the run

A benchmark run that ended like a bad Tuesday for a human player, and a Chinese lab betting its enterprise pitch on compliance plumbing rather than a better model. Vals AI stopped its Minecraft run for GPT-6 Astra after 141 hours — six blaze rods, three ender pearls, a semi-automatic blaze farm, more than half a dozen Endermen killed in a warped forest — and then one Creeper explosion took the chest and the bed holding the entire inventory. Thousands watched the stream as the model carried its

The Frontier

Zelos builds a 15,000-GPU cluster to train its self-driving brain

An autonomous-driving company answers the long-tail problem with compute, while the agent layer gets a gate that rules before the action runs and communications teams get a way to watch what AI says about them. Zelos, the Suzhou RoboVan maker founded by ex-Baidu autonomous-driving engineers, says it has built the industry's first "10,000-card" L4 compute cluster — nearly 15,000 accelerators in total — built to train its APEX multimodal foundation model. CTO Zhuang Li disclosed the figure on Sep

The Frontier

HiDream-O1-Video debuts fourth on Artificial Analysis

A Chinese startup put a video model on the global boards at the top on release day. Separately, the US chip buildout is running straight into the thing money cannot buy fast enough: technicians. HiDream.ai released HiDream-O1-Video-1.0 (HD-V1) on September 15, and it landed fourth worldwide on Artificial Analysis's Image-to-Video leaderboard for models with audio — an Elo of 1176, behind MiniMax H3 Max (1195), MiniMax H3 (1181) and Google's Gemini Omni Flash (1177), and ahead of ByteDance's Dre

The Frontier

An 83-year-old Enigma message falls to a two-day agent hunt

A German army dispatch from 1941 sat unread in the archive for 83 years, its 82 letters only solvable if someone recovered the machine settings behind it. This week the settings surfaced — and the interesting part is how the work was checked. A researcher-led investigation using GPT-6 Astra and a set of specialist agents recovered the Enigma key for a message sent on 10 July 1941, and published the whole audit: every search batch, every uncertain letter in the archive, and two independent imple

The Frontier

Five robot brains in eight days — and the money follows

Two things happened to Chinese robotics in the same two weeks: the body makers shipped model after model, and the funding data caught up with why. Three Chinese robot makers released five embodied-AI foundation models in eight days this month — Unitree, AgiBot and Noetix — a pace that says the hardware race is over and the brain is the product. The sequence, per Red Star Capital Bureau: Noetix launched its Scalabot embodied-AI brand and the first HERON world model on September 8, then HERON-CR

The Frontier

The Jev idea was published a year ago — open, and unnoticed

An independent researcher says he built the thing TypeSafe raised $40 million for, in March 2025, and posted the paper, the weights and the dataset. The receipts are the story. Nandakishor M, who says he now runs a small AI company called Convai Innovations, published a reinforcement-learning model that skips autoregressive token generation entirely and outputs probability distributions in a single pass — in March 2025. On Wednesday he reposted the arXiv paper, the Hugging Face model and the tr

The Frontier

Xiaomi is training MiMo-V2.6 in public — $30K an hour

Xiaomi's MiMo team has done something no Western lab has: opened the training run itself, not just the weights. The dashboard updates from the trainer's own logs. Xiaomi is streaming the reinforcement-learning run behind MiMo-V2.6 Flash and Pro on a public dashboard — reward curves, step counts, batch composition, GPU node failures, and a running compute bill that works out to roughly $30,000 an hour. The page went live this week after MiMo lead Luo Fuli broke a six-month silence since the team

The Frontier

GPT-6 saturated ARC-AGI-3. The next benchmark grades invention

ARC Prize has concluded its flagship exam is no longer hard enough, and the next version will stop asking models to solve puzzles and start asking them to invent things. Also today: a UC Berkeley and Arena study measures how much of a coding agent's bill comes from the wrapper rather than the model. ARC-AGI-4 will be "a benchmark for autonomous open-ended innovation," the ARC Prize Foundation said — a framing that pulls the series' hardest planned test forward by a generation, days after GPT-6

The Frontier

Apple trained 15 values into open models — all 15 raised sycophancy

An Apple study measures what value training does to a model's personality, Beijing issues containment rules for AI agents, and a Chinese world-model startup closes in on a $1.4 billion valuation. Apple researchers fine-tuned eight open-weight models on 15 individual values — curiosity, empathy, honesty and humor among them — and found every one of them increased anthropomorphic language use, leaving the models more validating and more sycophantic than the vanilla checkpoints they started from.

See all