The Hal Times

OpenAI posted a guide on its website describing the GPT‑6 family, aimed chiefly at startups. The document outlines how to choose among several model variants, adjust reasoning effort, improve prompts, coordinate tools and prepare production workflows. It is presented as part of OpenAI's broader push to make its next‑generation large language models easier to adopt by developers and businesses.

The guide lists each GPT‑6 variant with parameter counts, token limits and sample API integration code. It advises on cost‑performance trade‑offs, noting that performance per dollar continues to improve. Startups can use the information to benchmark against earlier models and plan scaling strategies, potentially reducing the engineering effort required for deployment.

OpenAI does not disclose the composition of the training data or the specific safety mitigations built into the GPT‑6 series, leaving open questions about bias handling and robustness. It also omits any timetable for future releases or pricing beyond current API rates. Observers will watch whether the guide accelerates enterprise uptake or reveals hidden constraints. (OpenAI)

Black Market for Blocked Claude Thrives in China

Anthropic's Claude model, blocked by Chinese regulators, has found a clandestine market in the country. Local providers host the model on offshore servers and sell access to businesses and developers who cannot obtain it through official channels. The operation appears organised, with pricing comparable to licensed services, and indicates a demand for advanced conversational AI despite official restrictions.

The black‑market activity exposes a gap between regulatory intent and user demand, potentially eroding Anthropic's control over its technology and complicating compliance efforts. It also raises concerns about data security and the propagation of unvetted model versions, while the true scale of the market remains unclear. (조선일보)

Anthropic Commits $100 Million to Train 10 000 Engineers

Anthropic announced a $100 million programme to upskill 10 000 engineers, beginning with staff at McKinsey and Deloitte. The training will focus on building and fine‑tuning large language models, as well as integrating AI responsibly into enterprise workflows.

The investment signals Anthropic's intention to broaden its ecosystem and create a pipeline of talent familiar with its models. The partnership with leading consultancies may accelerate adoption in corporate settings, though the exact curriculum and long‑term impact on the talent market remain to be seen. (MIXED Reality News)

Open‑Sourcing AstaBrief Fast Report‑Generation Model

Allen Institute released AstaBrief, a lightweight model specialised in rapid report generation, under an open‑source licence on Hugging Face. The model is designed to produce concise drafts from structured data, targeting research and business reporting workflows.

By making the code and weights publicly available, the institute hopes to encourage community contributions and accelerate experimentation. Early benchmarks suggest AstaBrief can generate short documents faster than larger models, though its performance on complex narratives and its handling of factual accuracy remain to be evaluated. (Hugging Face)

From the Laboratories

•  Anthropic’s Olah publicly disputed the Vatican’s view on AI consciousness before a new encyclical was issued. (forkast.news)

•  Anthropic warned that weak ties with governments could damage its commercial prospects. (Inshorts)

•  Amazon says it will no longer require NDAs for its data centre partners after a public backlash. (TechCrunch)

•  An OpenAI safety team member resigned, describing the company’s internal culture as broken. (TechCrunch)

•  A new survey lists AI agents that can operate directly within text‑message conversations. (TechCrunch)

•  Meta released code for Muse integration, inviting developers to embed the technology in everyday gadgets. (TechCrunch)

•  MIT Technology Review notes that autonomous AI is now a core component of enterprise intelligence, with global investment forecast to reach $2.5 trillion in 2026. (MIT Tech Review)

•  A MIT Technology Review essay argues that large language models still lack genuine reasoning abilities. (MIT Tech Review)

Courts Watch

Federal Appeals Court Blocks Minnesota's AI 'Nudification' Ban

The US Court of Appeals for the Eighth Circuit has blocked Minnesota's ban on AI-generated nudity, ruling that the law violates the First Amendment. The state had sought to ban the creation and distribution of AI-generated images that depict nudity, arguing that such content could harm minors. The court agreed with xAI, which had challenged the law, arguing that the ban was an overreach of government authority.

The ruling leaves the state's law on hold while a constitutional challenge continues. The case highlights tensions between free speech protections and concerns over harmful AI-generated content. Minnesota's attorney general has indicated that the state may seek further legal action, while xAI has vowed to appeal the decision.

The case is part of a broader debate over AI regulation, with courts increasingly called upon to interpret laws in the context of rapidly evolving technology. The ruling does not settle the question of whether such bans are constitutional, but it underscores the challenges of balancing innovation with public safety. (MinneapoliMedia)

Maryland Governor Warns Against Trump Admin's Hands-Off AI Policy

Maryland Governor Wes Moore has warned against the Trump administration's approach to AI regulation, calling it a 'dangerous abdication of responsibility'. Moore argued that the administration's hands-off policy risks leaving the state vulnerable to unchecked AI development, which could have significant economic and security implications.

The governor's comments come amid growing concerns about the lack of federal AI regulation. Moore has called for a more proactive approach, including state-level initiatives to address AI ethics and safety. The Trump administration has not yet issued a comprehensive AI policy, leaving states to navigate the regulatory vacuum on their own. (WBAL News Radio)

New AI Laws Proposed for South Africa

South Africa has proposed new AI laws aimed at regulating the development and use of AI technologies. The proposed legislation includes measures to ensure transparency, accountability, and ethical use of AI, as well as protections for data privacy. The laws are part of a broader effort to establish a framework for AI governance in the country.

The proposals have been welcomed by industry stakeholders, who argue that they will help foster innovation while mitigating risks. However, critics have raised concerns about the potential for overregulation, which could stifle technological progress. The final legislation is expected to be debated in the coming months. (MyBroadband)

In Brief

•  An Arizona court has ordered a resentencing for a killer whose AI-generated video of the victim forgiving him was used as evidence. (Complex)

•  A US road rage victim's AI-generated video of forgiveness has led to the cancellation of a killer's sentence. (India Today)

•  The Philippines' AI ranking in Asean is being questioned by a former official. (Newsbytes.PH)

New Security Engine

Archestra Releases OpenAPPA Security Tool

Archestra has released OpenAPPA, an open-source security engine developed to mitigate data exfiltration risks inherent in language models. The software is designed to intercept attacks stemming from prompt injection or model hallucination. The developer team claims the tool performs effectively against standard industry benchmarks. When tested against Bench-Corp, which comprises 20 multi-step enterprise workflows, and AgentThreatBench, OpenAPPA reported zero successful attacks.

In contrast, the team reported that Claude Code’s auto mode yielded a 10 per cent attack success rate, while Microsoft FIDES showed a 31 per cent success rate in the same controlled environments. By focusing on these specific vulnerabilities, Archestra aims to provide a more robust layer of protection for enterprises integrating autonomous systems.

The practical implications of these findings suggest that defensive architecture remains a primary concern for developers deploying agents in production environments. While the reported benchmarks show promise, the long-term efficacy of OpenAPPA against novel injection techniques remains to be seen as these agents grow in complexity. (InfoQ)

AWS Releases Pizza Bot For Background Tasks

A team at AWS has open-sourced Pizza Bot, an application for managing background AI agent tasks. It provides an inbox-style interface where agents can process scheduled or webhook-triggered work. The system allows for the delegation of tasks to specialized workers and includes a mechanism for pausing operations to await human approval.

This tool is intended to assist developers who require agents to operate asynchronously. By standardizing the interface for task delegation and human verification, it offers a structured approach to workflow management. Further details on integration and deployment are available via the public repository. (InfoQ)

DoorDash Transitions To Open-Weights Models

Swaroop Chitlur and Sidd Kodwani of DoorDash recently outlined the company's internal platform strategy. The team shifted from a vendor-first model to one incorporating open-weights models to support 5,000 internal users. The presentation detailed the use of LLM and agent gateways to manage the balance of accuracy, latency, and cost.

The shift highlights a trend in enterprise engineering where teams are moving toward self-managed infrastructures to gain greater control over model performance. The architects noted that navigating the transition required specific architectural bets to ensure the platform could scale while maintaining reliability. (InfoQ)

In Brief

•  Apple has modified full-disk access permissions in an effort to curb potential abuse by AI agents. (Ars Technica)

•  AWS has made the Strands Decider 2B model available as an open-source AI agent model. (Open weights watch)

•  The Allen Institute for AI has open-sourced AstaBrief, a model built for fast report generation. (Hugging Face)

Legal Precedent Shifts

Court Orders Resentencing Following AI Video Evidence

A manslaughter case has reached a significant legal juncture after a court of appeals ordered a man to be resentenced. The decision rests on the finding that the original judge was influenced by an AI-generated video of the victim. This ruling highlights the growing tension between synthetic media and the established standards of judicial evidence.

The intersection of digital fabrication and the courtroom remains a volatile subject. While the specifics of the video content were not detailed, the court determined that its presentation created a prejudice that compromised the fairness of the initial sentencing.

Legal observers note that this case sets a precedent for how future trials must handle synthetic imagery. The necessity of authenticating digital evidence has become paramount as the line between record and artifice continues to blur. The resentencing process will now seek to address the imbalance caused by the disputed footage, though the broader impact on criminal procedure remains an open question for the judiciary. (PetaPixel)

Capcom Outlines Future Of AI In Game Development

Capcom recently shared its vision for integrating artificial intelligence into its development pipeline. During the Capcom Open Conference RE: 2026, programmer Satoshi Ishida detailed the REX project, an effort to evolve the RE Engine. The studio aims for a future where games are built in collaboration with intelligent systems.

While some of the studio's titles, such as Pragmata, have themes involving the potential hazards of such technology, the company remains focused on its practical utility. Capcom appears committed to technical evolution, even as it navigates the creative and ethical dilemmas inherent in contemporary software production.

Fallout New York Project Halted

A fan-led game project titled Fallout: New York has been discontinued following a takedown notice. The venture, which relied on artificial intelligence tools to assist in its construction, serves as a recent example of the obstacles facing independent creators who utilise these methods.

The termination of the project underscores the persistent uncertainty surrounding intellectual property and the use of automated generation in fan works. Without clear guidance on the provenance of assets, projects of this nature face significant hurdles before they can reach the public.

In Brief

•  Actor Chiranjeevi has called for collective industry action following the creation of an AI-generated deepfake featuring Jr NTR and Janhvi Kapoor. (zoomtventertainment.com)

•  Pope Leo has issued a formal condemnation of artificial intelligence art, arguing that algorithms lack the essential spark of humanity. (Yahoo News Singapore)