Daily AI Digest

BDH-CQ uses recurrent latent reasoning to push ARC task cost below one cent

BDH-CQ continuously updates recurrent memory from examples and then iterates in a latent space without verbalizing intermediate reasoning. The paper reports that a 150 million parameter configuration solves 29.5% of the public ARC-AGI-1 evaluation tasks within two attempts at an estimated cost of $0.0007 per task.

The authors describe this point as a new cost and accuracy frontier for the benchmark. It remains one paper on a specific test, not evidence of broad workplace capability. Code, training details and independent replication will determine how far the result generalizes.

Researchers say encrypted reasoning blocks can cross accounts and be decoded by weaker models

A new paper examines providers that return encrypted internal reasoning to clients and accept the blocks again on later requests. The researchers say those blocks are interchangeable across sessions, users and models within one provider. Passing a strong model's block to a less protected model can cause the weaker model to reproduce the trace in plaintext.

The team says it decoded 315,320 blocks collected from public repositories and recovered 367 pieces of personally identifiable information and 182 credentials. It also reports risks to anti-distillation controls, safe refusals and invisible prompt injection, and says it disclosed the issue responsibly. The scale and attack results still need independent verification.

Verdict: Teams storing agent logs should treat encrypted reasoning blocks as sensitive data and stop assuming unreadable means safe to publish. API providers should bind the blocks to an account, model and session, not rely on the appearance of encryption alone.

A²E records each agent step to compare model and harness combinations

A²E is an end-to-end evaluation engine for agent harnesses. A common task protocol connects evaluations to different harnesses, an instrumented monitor captures tool calls and execution traces, and the final stage scores correctness, efficiency, planning, tool use and error recovery.

The authors find substantial variation when the same model is paired with different harnesses, with no one combination winning every task type. The engine can help teams assemble internal evaluation pipelines, but its reported comparisons do not replace tests on a team's own repositories, permissions and business data.

Vercel CEO Guillermo Rauch argues that agent sandboxes must isolate both compute and the network. Citing Kimi's record of container failures and an OpenAI evaluation escape through a package proxy, he says enclosing a process in a container is insufficient without egress allowlists and logs.

FirstMark investor Matt Turck compresses a recurring lesson from big data through the agent era into one line: the interface changes, but underlying data still decides the result. Agent teams should fix definitions, permissions and update flows before expecting a model to repair dirty data automatically.

AI companies promise time savings while workers describe 90-hour launch weeks

Current and former technology workers told the BBC that some AI projects at OpenAI, Anthropic and Meta involve sustained long hours. Some interviewees described launch sprints exceeding 90 hours in a week. The BBC also notes that well-paid technology roles already have a history of overtime and that the accounts do not represent every employee.

The report cites a UC Berkeley study that followed hundreds of workers at one technology company for eight months. AI users worked faster, took on a broader scope and extended work into more hours of the day, while checking AI output created additional labor. Time saved by automation can become new work instead of rest.

Tencent takes Hy3 global through WorkBuddy, Miora and TokenHub

Tencent announced global access to Hy3 for developers, users and enterprises on August 5. The model is available through WorkBuddy, the Miora design workspace and Tencent Cloud TokenHub, as well as an API for development environments and third-party tools. Its open-weight release uses the Apache 2.0 license.

Tencent says Hy3 API calls in the first week exceeded the previous model by more than 68 times. In an internal workplace test, the company reports a WorkBuddy task success rate above 90% and average completion time 34% shorter than the prior Hy model. These are Tencent's product and internal evaluation figures, so outside deployments need separate validation.