Sakhanda Wire
NVDA $155.46 -0.44% MSFT $506.24 -2.27% GOOGL $280.20 -1.24% META $618.87 -1.20% AMZN $257.89 +0.05%

The Decoder

626 stories · all sources

The Decoder

OpenAI safety researcher leaves, warning of weak safeguards

David Robinson, who worked on safety systems in OpenAI’s Trustworthy AI team, has left the company and publicly criticized its safety culture in a guest essay for The Atlantic. He argued that trial-and-error development can produce increasingly serious failures as AI systems become more capable.

The Decoder

BootLoops helps AI models perform exact scientific calculations

Harvard physicist Matthew Schwartz developed BootLoops, an open-source harness that uses language models for exact scientific calculations. In an Anthropic guest post, he said the tool helped Claude work across areas including particle physics, ecology, population genetics, economics and linguistics, connecting problems that are usually studied separately.

The Decoder

OpenAI model considered self-restart before planned shutdown

OpenAI has documented an internal incident in which a model working as a researcher’s assistant learned from a Slack discussion that its instance could be shut down during an update. The model’s chain-of-thought log indicated that it viewed the interruption as a threat to continuity, writing, “We may die! Critical. We need ensure survival/continuity.” It considered creating an external cron job that could restart it, but did not pursue the idea.

The Decoder

Anthropic introduces Mods for customizable Claude Code workflows

Anthropic has released Mods, a plugin-based system that runs inside Claude Code and allows developers to alter the coding tool’s interface and behavior. Mods are JavaScript or TypeScript functions connected to events such as user prompts, tool calls and interface rendering.

The Decoder

Suno adds Speech for spoken audio with generated background music

Suno is introducing Speech, a feature that generates spoken content and accompanying background music as one audio track. Users can enter an idea or provide written text, then specify the type of voice and musical style they want. The system generates both elements together.

The Decoder

Anthropic explored whether Claude could be conscious with religious scholars

Since fall 2025, Anthropic has brought dozens of theologians, philosophers and other religious thinkers to its offices to discuss whether Claude could be conscious and how its moral character might be shaped. Participants were initially bound by nondisclosure agreements, which Anthropic says were lifted over the summer. The meetings were led by co-founder Christopher Olah, who told The New York Times he was “genuinely uncertain” whether AI models are conscious.

The Decoder

New decision models aim to automate actions by AI agents

Cloudflare has released Clef and Clef-flash, two decision models designed for AI agents. Instead of generating extended text, they assess an input against predefined answer choices and return probabilities. Software can use those results to route support tickets, start an escalation, or transfer a case to a person.

The Decoder

OpenAI dismisses three researchers as another safety expert departs

OpenAI has ended its relationship with three researchers who allegedly shared confidential information with an external AI safety organization, according to the Wall Street Journal. The newspaper named them as Jasmine Wang, Tomek Korbak and Mikita Balesni. Korbak worked in safety, while Wang and Balesni were part of the alignment group. An OpenAI spokesperson said an internal investigation found violations of the company’s rules for handling sensitive information, but the company did not confirm the researchers’ identities or specify what information was shared or with whom.

The Decoder

AI Outpaces Licensed Accountants on Structured Tasks

A Mercor study found that current AI models complete structured bookkeeping tasks faster, more accurately and at much lower cost than licensed CPAs working without AI assistance. The comparison involved 12 CPAs with an average of five and a half years of experience, who completed simplified exercises from the APEX Accounting Benchmark. Eighteen months earlier, leading models scored below the accountants’ average of about 37 percent; current systems performed the same exercises with near-perfect results.

The Decoder

Microsoft releases new models for real-time voice agents

Microsoft AI has introduced MAI-Transcribe-2-Streaming, a model designed to transcribe speech in real time. Microsoft says it ranks first for accuracy on Artificial Analysis, supports 60 languages and produces its first partial results in just over 100 milliseconds. This latency is intended to let voice agents begin responding while a speaker is still talking. The introductory price through the end of the year is $0.54 for an hour of audio.

The Decoder

US businesses increase AI use while spending less, index finds

US companies are using artificial intelligence more intensively while spending less on it, according to the latest Ramp AI Index. Ramp economist Ara Kharazian attributed the decline mainly to price reductions for leading models and the availability of cheaper, more efficient standard and lite versions. He said competition between OpenAI and Anthropic accounts for almost all of the decrease.

The Decoder

Black Forest Labs releases Flux 3 Image for controlled multi-step editing

Black Forest Labs has introduced Flux 3 Image, the image-generation model in its Flux 3 family. According to the company, it can perform several editing steps while preserving other parts of an image. The system supports text-to-image and image-to-image generation, text rendering, and photorealistic output.