Daily Brief - SiliconNoon (Page 115)

The Guardrails

South Korea's national AI race faces benchmark-rigging claims

Seoul's sovereign AI project — a government-funded contest to crown one "national representative" foundation model — heads into its second-round verdict under a cloud: foreign firms allegedly offered to juice the entrants' benchmark scores. Allegations that overseas companies offered to inflate benchmark scores for South Korea's government-backed AI foundation model race surfaced days before its second-round elimination. According to industry sources reported by The Hankyoreh, US-based training

The Guardrails

OpenAI dissolves the team built to catch catastrophic AI risks

OpenAI has quietly dismantled the unit whose entire job was deciding whether its own models could cause catastrophic harm — the clearest organizational sign yet of where safety sits as the lab races an IPO clock. OpenAI dissolved its "Preparedness" team at the end of July, the Financial Times reports, parceling its biological and cyber risk evaluation work out to existing groups. The unit ran the framework that last week forced the company to flag its upcoming Astra model as potentially "Critic

The Frontier

LittleLearner: LLMs can't outgrow their training data

A research team from MPI for Intelligent Systems, ELLIS Tübingen, and ETH Zürich trained language models from scratch on nothing but U.S. elementary-school material — and found that no amount of post-training or clever prompting pushes them past the fifth-grade ceiling their pretraining data set. Researchers built LittleLearner, three LLMs (0.6B, 1.3B, and 5B parameters) trained from scratch on LittleCurriculum — an 88B-token corpus distilled from FineWeb-Edu through a five-stage filter aligned

The Guardrails

Anthropic ran 133M contractor chats with bio-weapons filters off

A safety filter nobody noticed was missing, at the lab whose CEO calls bio-weapons risk the scariest thing in AI. Anthropic disclosed that its biological-weapons blocking classifiers were offline for nearly a year — from May 2025 through April 2026 — leaving roughly 133 million model exchanges from about 50,000 human-feedback contractors completely unfiltered. The admission appears in the company's August 2026 redacted risk report. The failure wasn't a jailbreak that slipped past a safeguard; t

The Arena

Huawei's open-source WorkSwarm brings swarm agents to HarmonyOS PC

The multi-agent trend crossed a new threshold this week: openJiuwen, the open-source agent platform built by Huawei's 2012 Labs, Huawei Cloud and the terminal team, launched WorkSwarm — a swarm of office agents that collaborate as a team, not a single assistant — and it's the first of its kind on the HarmonyOS PC app store. Huawei-backed openJiuwen upgraded its JiuwenSwarm engine into WorkSwarm, a full office-agent suite that ships on the HarmonyOS PC app store with Windows and Mac support. The

The Frontier

Gowers: LLMs solve math's biggest problems with counterexamples

A Fields Medalist's honest read on the LLM math boom, and a new benchmarking tool that tests models against your own data instead of someone else's. Fields Medalist Timothy Gowers says the most famous mathematics problems LLMs have solved so far came almost all as counterexamples, not proofs. Writing days after OpenAI announced it had cracked ten major open problems — including the first construction of a non-sofic group and a superexponential lower bound for multicolor Ramsey numbers — Gowers

The Frontier

Pathway raises $30M at $500M for post-Transformer AI models

Two stories about AI economics this morning: one lab just landed a hefty seed valuation for betting that reasoning doesn't have to be expensive, and a respected benchmarking outfit launched a tool for testing whether models actually fit your work. Pathway, an AI research lab betting that reasoning doesn't have to be expensive, has raised a $30 million seed round at a $500 million valuation to scale its "post-Transformer" BDH architecture. The company's flagship result so far: a 150-million-para

The Everyday

One in five US workers now hands tasks to AI, not colleagues

A new representative survey from Epoch AI and polling firm Ipsos quantifies how far AI has crept into everyday work: one in five employed Americans now hands at least one task to a machine that used to go to a person. One in five employed US adults now delegates at least one work task to AI that previously went to a coworker or outside contractor, according to a survey of 1,106 workers by Epoch AI and Ipsos. The poll, fielded July 10–19, asked about ten common knowledge-work tasks drawn from US

The Arena

Alibaba's Qianwen Office adds GLM-5.3 and DeepSeek V4 Pro

Alibaba just turned its flagship office agent into a showroom for its rivals' models — a move no Western office suite has made yet. Alibaba's Qianwen Office — the enterprise agent suite that opened public beta earlier this month — is the first product to offer Zhipu's GLM-5.3 and DeepSeek's V4 Pro directly, selectable today from the "frontier models" tier on the app's homepage. The two additions sit alongside Alibaba's own Qwen3.8, which shipped as the launch flagship when the product debuted i

The Frontier

AI agent ports 250k-line weather sim to GPU with 5.1x speedup

A new arXiv paper shows an AI agent modernizing one of the hardest kinds of legacy code — a 250,000-line Fortran weather model — and the validation-first workflow that kept the port trustworthy. Researchers used an AI agent to port CReSS, a legacy Fortran weather simulation code with more than 250,000 lines, to run on GPUs — and the ported version ran a real typhoon simulation 5.1x faster while staying numerically faithful to the original. The workflow, described in a paper submitted to the Ag

The Everyday

Why China is far more optimistic about AI than the US

Two countries, two AI publics. A Bloomberg weekend essay from Grace Shao puts hard numbers on the divide shaping everything from product strategy to regulation — and the gap isn't about who uses AI more. Chinese citizens are dramatically more optimistic about AI than Americans, and the difference comes down to who expects to benefit, not who sees more risk. In Stanford's AI Index, 84 percent of Chinese respondents said they're excited by AI — the highest share of any country surveyed — while in

The Guardrails

Amodei: open weights are 'nowhere near a sufficient solution'

Anthropic's CEO went where he almost never goes — a long X thread — to make the case for regulation in the middle of the industry's loudest argument about open weights. Dario Amodei took to X Saturday to defend AI regulation and mandatory frontier-model testing, arguing that open-weight models are "nowhere near a sufficient solution" to AI's concentration problem. The exchange began when a16z's Gavin Baker and Anthropic co-founder Sholto Douglas sparred over claims about Amodei's positions that

See all