DeepSeek Cuts API Prices by 75% — the Exact Opposite of Competitors' Price Hikes
DeepSeek has announced a 75% cut to its API prices. While other major AI labs have been raising prices two- to threefold across the board, DeepSeek is the only one moving in the opposite direction.
This pricing strategy appears to reflect DeepSeek's confidence in its cost-efficient model development. That said, the intensifying price competition is presumably driven by rising inference costs and the need to recoup infrastructure investments. While users gain more options, attention is focusing on how each company balances profitability with sustainability.
AI Coding Company Cognition Raises $1 Billion at $26 Billion Valuation
AI coding startup Cognition has raised $1 billion at a $26 billion valuation, according to Bloomberg.
Cognition is the company behind the AI programming assistant Devin and other products, and it has reached a massive valuation in a short period on the back of rapid growth in the AI coding space. The AI coding tool market is crowded with players including Devin, Cursor, and Windsurf, and the size of this raise is a testament to the market's expectations. That said, when it comes to the company's actual revenue base and product differentiation, we will have to wait and see how things unfold.
Hiring AI Rejects Black and Asian Applicants at Higher Rates — Bias Concerns Resurface
AI hiring algorithms have been found to reject Black and Asian job seekers at higher rates than other groups, according to a report by The Register.
AI hiring tools are already deployed at many companies, and it has once again been pointed out that bias inherent in training data may be influencing actual hiring decisions. Discriminatory outcomes have reportedly been confirmed at the resume screening and online assessment stages in particular.
Regulatory debates over AI fairness are progressing in various countries, but practical countermeasures are still considered insufficient. As AI use in hiring expands, ensuring transparency and establishing regular audit mechanisms have become urgent priorities.
AI Coding Agents Autonomously Install Packages That "Nobody Maintains"
A security risk has been flagged in which AI coding agents autonomously install packages with no maintainer or unknown ownership, according to a report by The New Stack.
When automating coding tasks, AI agents install the packages they need at their own discretion, creating the danger of malicious or abandoned packages slipping in along the way. Traditional dependency management has assumed human review, but as AI agents grow more autonomous, that assumption is crumbling.
Proposed countermeasures under discussion include sandboxing AI agents' install operations and introducing trust scoring for packages.
Open Source Developers Burning Out from AI-Driven Bug Reports — Heise Commentary
German technology media Heise reports that open source developers are being pushed into burnout as they deal with AI-related bugs.
With the spread of AI coding tools, a flood of AI-generated contributions is pouring into open source projects. Many of these look correct on the surface but actually contain subtle bugs, and the burden on reviewers has reportedly surged.
Maintainers buried in handling low-quality AI-generated patches are engaged in volunteer work, yet they are effectively being forced into nearly full-time labor. This raises fundamental questions about the sustainability of open source communities.
Qwen 3.5 35B Runs on a $300 Laptop — The Evolution of Local LLMs
A user report from Reddit's LocalLLaMA community is drawing attention: running inference on the Qwen 3.5 35B parameter model at 10.33 tokens per second on a $300 laptop.
Running LLMs locally is becoming increasingly important for privacy, lower API costs, and offline use. Getting a 35-billion-parameter-class model to run at practical speeds on inexpensive hardware is an achievement that greatly expands the possibilities of local AI.
That said, details of the execution environment, the model's quantization settings, and actual output quality await independent verification for reproducibility.
Warp Announces Open Source Development Workflow Using GPT-5.5
OpenAI's official blog featured how Warp, a terminal app, is using GPT-5.5 and OpenAI models to enable coding agents that work together across local, cloud, and open source development workflows.
Warp takes an approach of integrating AI into developers' everyday terminal operations, and this release reportedly focuses in particular on agent-to-agent collaboration in open source development. While this could take development workflow automation a step further, the difficulty of control and debugging amid increasingly complex inter-agent coordination is likely to draw criticism as well.
A 103-Billion-Token Usenet Corpus — Training Data with Zero AI Contamination
A Reddit user has built and released a 103-billion-token text corpus based on Usenet data from 1980 to 2013.
The corpus's biggest feature is "zero AI contamination." Every piece of text comes from an era before LLMs existed, so it contains no AI-generated text bias at all. As a clean dataset for fine-tuning, it is drawing strong interest from the local model development community.
With training data quality directly tied to model performance — and amid growing concern over recursive training on AI-generated text (model collapse) — the value of "human-only" corpora like this is likely to keep rising.