The Hal Times

Anthropic is preparing for an initial public offering with a valuation strategy tied to an ambitious 190 to 200 billion dollar revenue forecast for the year 2028. This figure serves as the primary benchmark for investors evaluating the firm's trajectory. While the company has secured significant interest from major partners, including Amazon, the reliance on such long-term projections highlights the speculative nature of current laboratory financing.

Elon Musk, who previously expressed reservations regarding the company's capabilities, has admitted to underestimating Anthropic. This shift in sentiment carries weight for Amazon shareholders, who maintain a vested interest in the laboratory's output. Despite this external validation, the path to the 2028 revenue goal remains unproven.

The firm also continues to signal caution regarding the wider industry. Anthropic has issued public warnings concerning the mounting risks associated with the proliferation of artificial intelligence. It is unclear how these stated safety concerns will reconcile with the commercial pressures of a public company tasked with meeting aggressive growth targets set for the end of the decade. (The Business Standard; The Globe and Mail)

OpenAI Previews Ultrafast Mode For GPT-5.6

OpenAI has introduced a new service tier for its API called Ultrafast, designed to accelerate the performance of GPT-5.6 Sol. Powered by hardware from Cerebras, the service reportedly delivers output speeds of up to 750 tokens per second, which the company claims is 14 times faster than standard operations.

The release includes updated documentation for builders, focusing on cost-efficient agent workflows and new capabilities within the Responses API. These tools are intended to assist startups in selecting models more effectively, though the specific long-term impact on operational overhead remains a subject for further evaluation. (OpenAI)

Meta Releases Glimmer Model

Meta has released Glimmer, an open-weight artificial intelligence model that allows users to operate the software on their own hardware. This approach contrasts with the company’s Muse Spark model, which remains accessible only through restricted APIs.

In a concurrent statement, Mark Zuckerberg argued that artificial intelligence should be widely available rather than concentrated within a few private laboratories. Whether this philosophical shift translates to broader adoption, or simply highlights the divide between open-weight distributions and proprietary systems, continues to be a point of debate among industry observers. (TechCrunch)

From the Laboratories

•  Google is permitting users to remove visible watermarks from AI-generated content, though invisible benchmarks will remain in place. (TechCrunch)

•  Google DeepMind has officially introduced the Gemini 3.7 Flash model. (Google DeepMind)

•  French startup Kog is investigating methods to improve inference efficiency on GPU hardware. (TechCrunch)

•  Rising natural gas prices present a financial risk to hyperscalers relying on the fuel for AI data centres. (TechCrunch)

EU Transparency Rules

Anthropic Adds Hidden AI Markers To Claude Outputs

Anthropic has begun embedding hidden markers into the outputs generated by its Claude models. This technical adjustment follows the implementation of new European Union transparency requirements. These rules mandate that companies provide clear indicators when content is produced by artificial intelligence, ensuring that users can distinguish between human and machine-generated information.

The invisible watermarking process does not alter the text visible to the user. Instead, it embeds data that remains detectable to specialized analysis tools. This approach aims to satisfy regulatory obligations without interrupting the natural flow of human-computer interaction. The firm has prioritised compliance as part of its ongoing integration into the European digital market.

What remains unclear is the efficacy of these markers against adversarial attempts to strip them from outputs. While the initiative addresses immediate policy demands, technical experts observe that the permanence of such markers depends on the development of robust detection standards across the industry. For now, the move marks a standardisation of transparency practices for the company. (Infomance)

Massachusetts Court Clarifies AI Work Product Protection

A Massachusetts court ruled that documents generated by artificial intelligence lack work product protection unless they are produced under the direct supervision of an attorney. The decision establishes that the mere use of an automated tool does not grant legal immunity to the resulting files.

The court emphasised that protection hinges on the human involvement and strategic direction provided during the drafting process. If an attorney does not explicitly guide the creation or vetting of the machine-generated material, it may be subject to discovery in legal proceedings. This ruling clarifies a point of concern for legal professionals adopting new tools. (londoninsider.co.uk)

Sturgeon Lake Cree Nation Challenges Data Centre Project

The Sturgeon Lake Cree Nation in Alberta has secured a victory in an initial legal challenge against a proposed large-scale artificial intelligence data centre. The nation raised concerns regarding the environmental and social impacts of the project, which they described as massive and mind-boggling.

This ruling represents a significant development in the oversight of infrastructure projects tied to the artificial intelligence sector. While the court has sided with the First Nation in this particular battle, the future of the development remains subject to further regulatory and legal scrutiny regarding land use and community impact. (APTN News)

In Brief

•  A South Korean artificial intelligence chip tycoon has filed an appeal against a record US$666m divorce settlement. (NST Online)

New Local Model

Meta Releases Muse Glimmer Open Weights Model

Meta AI Research has published Muse Glimmer, an open-weight model with 30 billion parameters. It operates under the Apache 2.0 license and targets local execution on consumer graphics processing units. By removing the requirement for cloud-based APIs, the model aims to support autonomous agent tasks and complex automation workflows directly on a user machine. It utilizes a multi-stage training method intended to maintain performance while remaining efficient enough for on-device deployment.

The model supports multimodal inputs, which expands its utility in coding and technical automation. This release provides an alternative to proprietary cloud services for developers who require data sovereignty or offline capabilities. Whether the model achieves performance parity with larger, cloud-hosted systems remains to be seen, as does its specific efficiency profile on various consumer hardware configurations. Documentation suggests it serves as a foundation for building self-contained agentic applications.

This development arrives alongside a broader shift in the sector, as competition increases among open-weight and closed models. Meta's release follows recent activity in the Chinese research landscape, including the introduction of GLM-5.3 by Z.ai, which also addresses complex coding requirements. As labs and firms refine these offerings, the architecture of agentic workflows is becoming a point of focus, with many practitioners debating the merits of large context windows against the precise, contextual engineering of smaller, targeted toolsets. (InfoQ)

Techniques For Improving Coding Agents

Software architects are reconsidering the value of large context windows in coding agents. Baruch Sadogursky and Patrick Debois argue that massive, noisy prompts often lead to failure. They propose an alternative strategy based on context engineering, which includes the use of versioned context artifacts and externalized memory banks. By moving away from stuffing prompts with excessive data, developers may improve the reliability of their workflows.

The approach relies on lazy-loaded skills and the use of LLM-as-a-judge evaluations to ensure quality. The goal is to transform raw documentation and markdown files into structured, actionable inputs for agents. This methodology prioritises precise, relevant data over volume, suggesting that a focused 300-token input often outperforms a bloated 100k-token prompt. (InfoQ)

Simplifying Tagging With Vector Embeddings

Simon Willison suggests a novel approach for tagging large collections of digital content. Instead of feeding an extensive list of existing tags into a large language model, which can be inefficient, he recommends generating new tags without reference to the current vocabulary. Once generated, these tags are compared against existing categories using vector embeddings.

This method avoids the complexity of long-context constraints and limits the potential for models to be overwhelmed by broad tag sets. It relies on the mathematical proximity of vectors to find suitable matches. This technique allows for automatic tagging of older archives without the need for manual curation. (Simon Willison)

In Brief

•  GitHub has updated its platform to allow software delivery workflows to run via agent apps without leaving the site. (GitHub)

•  Google is pursuing homomorphic encryption to allow for private AI computation without exposing raw user data. (Hacker News)

•  The Cosmograph app now offers an AI Model Atlas, visualising the relationships between various machine learning models as a 3D graph. (Hacker News)