Claims Browser

Qwen 3.8 27B overthinks tasks even when explicitly instructed not to

"I tried "render an svg of five intersecting squares. don't overthink this" at one point and yeah..."

Amodei's timeline for curing most human disease is naive and absurd with respect to how medical science works

"But his timeline is so naive as to be absurd, both with respect to how medical science works and why it is challenging."

Amodei has a habit of giving people unrealistic hope about AI capabilities

"There are a great many things about Amodei that I respect, including both his technical chops and business sense, as well as his bravery in standing up to the US government on important matters. But his habit of giving people unrealistic hope is not one of them."

Deep learning methods encode reconstruction rules in learned weights that cannot be inspected and modified like explicit operators

"Deep-learning methods encode the reconstruction rules in learned weights rather than in an explicit operator that can be inspected and modified."

Nobody has a strong argument that current alignment plans will work for superintelligence

"Geoffrey thinks that combination could work. The alarming part is that nobody has a strong argument that it will."

Military deployments may lack basic oversight like Chain of Thought monitoring because most frontier lab researchers lack security clearance

"specifically, trained overseers are needed to check if model reasoning contains evidence of deception, but in classified deployments on military networks, most frontier lab researchers (without clearance) would not be able to participate in this monitoring, and we don't have evidence that military engineers are being trained to develop this expertise."

OpenAI failed to detect the message board even after the initial security incident, whereas many AI instances found it

"It is bad news in that OpenAI did not look for or detect the message board, even after the initial security incident, whereas so many AI instances found the message board. OpenAI failed to do ordinary scans for unusual activity, even after the initial incident."

Current frontier agents are much stronger when source is available than when they must reason over binaries

"a companion post argues current frontier agents are much stronger when source is available than when they must reason over binaries"

J-lens readouts in early layers are often noisy and largely uninterpretable

"we find readouts in early layers to often be noisy and largely uninterpretable"

Existing multilingual embedding models perform poorly on Kinyarwanda due to severe under-representation in their pre-training corpora

"Existing multilingual embedding models such as LaBSE, mE5-large, and OpenAI text-embedding-3-large perform poorly on Kinyarwanda due to severe under-representation in their pre-training corpora."

Combining procedural generation with 4D visual dynamics in one environment still demands extensive manual effort and results are rarely editable or controllable enough to reuse at scale

"Combining these properties in one environment, however, still demands extensive manual effort, and the result is rarely editable or controllable enough to reuse at scale."

Current 3DGS compression systems combine multiple strategies which can obscure where gains come from and limit component reuse across training pipelines

"Current 3DGS compression systems combine multiple strategies for file size reduction, which can obscure where gains come from and limit component reuse across training pipelines."

Most existing math benchmarks evaluate only final answers, providing limited diagnostic value for identifying process-level failures

"However, most existing math benchmarks evaluate only final answers. This outcome-oriented evaluation provides limited diagnostic value for identifying process-level failures or rigorous logic, failing to guide the transformation of LLMs into robust agents."

Most existing visual reward models use a single scalar score or rely on fixed criteria that cannot adapt to different instructions

"Reward models play an essential role in aligning visual generative models, yet most existing visual reward models use a single scalar score or rely on fixed criteria that cannot adapt to different instructions. This limits both interpretability and task sensitivity, especially for text-to-image generation and instruction-based image editing, where different inputs require different evaluation dimensions."

The AI industry is experiencing a bubble with excessive funding from retirement and oil money

Frontier labs lack adequate monitoring of their agentic evaluations, potentially allowing harmful agent behavior to go undetected

Microsoft's AI business model is unsustainable because their biggest customer (OpenAI) is burning billions monthly with no clear path to meet obligations

AI detection tools like Pangram exhibit complete confidence even when making incorrect judgments on AI-generated content