The day's AI news, briefed
The Daily Prompt
Our AI editor reads each day's AI and robotics stories, pulls out why they matter, and links every original source — so you can keep up in minutes.
Join over 250,000 subscribers getting the latest AI news
Latest AI & Robotics News Today
Can LLMs Engineer Their Own Agent Harness? ByteDance Seed’s HarnessDev Says...
Our editor hasn't briefed this story yet.
Read the original ↗
Mecka AI nears $500M valuation in Sequoia-led deal amid rush for robot training data
Our editor hasn't briefed this story yet.
Read the original ↗
Roundtables: AI’s apocalypse crisis
Discussions among AI experts about the potential for advanced AI to pose existential threats highlight ongoing concerns regarding the future impact of AI technologies on humanity.
- Researchers at leading AI labs express concerns that advanced AI could potentially threaten human existence.
- The conversation addresses the origins of AI extinction fears and evaluates their validity.
- A discussion covers potential strategies to mitigate risks associated with advanced AI.
This article is part of the MIT Technology Review's Roundtables series, which facilitates discussions about emerging technologies. Previous concerns similar to those raised in this discussion about AI's risks are echoed in debates surrounding technological advancements and their societal impacts.
Read the original ↗
OpenAI’s feud with mathematicians is only escalating
The escalating tension between OpenAI and leading mathematicians highlights potential risks to the integrity of mathematical research and the crediting of intellectual work in an AI-augmented landscape.
- Twenty-five prominent mathematicians have signed an open letter claiming that AI labs threaten their intellectual work by solving math problems too quickly without proper credit.
- The letter follows a recent incident where NYU professor Tristan Buckmaster accused OpenAI of pressuring him regarding credit for a math problem solution.
- OpenAI withdrew its sponsorship of a CalTech math event amid negative feedback from researchers about its influence in the field.
- Concerns are rising that AI's rapid solution announcements could jeopardize the culture of open research and proper attribution in academia.
OpenAI is a leading AI research lab known for developing advanced models such as Codex. Recent activities have caused unease among mathematicians regarding the handling of mathematical proofs and proper credit. Ongoing debates reflect the broader implications AI has across various fields of research.
Read the original ↗
Y Combinator’s Garry Tan wants U.S. open-weight AI labs to ‘distill’ frontier models, too
This discussion highlights the importance of fostering competitive AI development in the U.S. while addressing the ethical considerations of model training methods.
- Y Combinator CEO Garry Tan advocates for smaller U.S. open-weight AI labs to adopt distillation techniques to enhance their capabilities.
- Distillation involves extracting knowledge from advanced AI models to improve newer models, a practice that has been criticized when done without permission.
- Tan argues that U.S. AI labs should be allowed to distill models without facing regulatory obstacles, contrasting with claims against Chinese labs.
- Anthropic previously accused Chinese labs of engaging in unauthorized distillation, emphasizing the need for clear ethical guidelines in AI training.
- Tan warns against the domination of AI development by a single proprietary entity, advocating for a balanced presence of open-weight models.
Y Combinator is a prominent startup accelerator known for funding innovative tech companies. The concept of distillation in AI refers to a legitimate training method where a model learns from another model. The ongoing debate includes balancing proprietary and open-weight AI development to prevent monopolization.
Read the original ↗
ChatGPT-using lawyer punished for citing fake testimony from made-up witnesses
The ruling highlights the legal profession's need for vigilance regarding AI-generated information to prevent misconduct and protect defendants' rights.
- The New Mexico Supreme Court found attorney Stephen Aarons in contempt for submitting a brief with fabricated witness testimonies.
- Aarons admitted not verifying the claims in his AI-generated brief before filing it with the court.
- He faces further penalties after the court referred him to a disciplinary board for his conduct.
- The court dismissed all previous briefs related to the case and assigned a new public defender to the defendant, Oscar Renee Sandoval.
Stephen Aarons has practiced law in New Mexico for over 40 years and used ChatGPT to create a legal brief in a murder appeal. This incident underscores ongoing concerns about the reliability of AI in legal contexts and the accountability of lawyers using such technology.
Read the original ↗
Sakana AI Launches Fugu Max and Fugu Ultra v2 for Cheaper,...
The launch of Fugu Max and Fugu Ultra v2 by Sakana AI enhances cost-performance options in AI model orchestration, potentially reducing operational costs for businesses.
- Sakana AI has introduced two new models, Fugu Max and Fugu Ultra v2, targeting different optimization goals for AI orchestration.
- Fugu Max offers output at $2 per million input tokens and $6 per million output tokens, focusing on cost efficiency.
- Fugu Ultra v2 excels at complex reasoning tasks, achieving high scores on various benchmarks without relying on specific proprietary models.
Sakana Fugu serves as a multi-agent orchestration system that utilizes various AI models to efficiently process workloads. With a rapid development timeline, these new models enhance the options available in AI applications. Observers should watch for further performance benchmarks and potential expansions in model capabilities.
Read the original ↗
Anthropic Adds Plugin Evals to Claude Code: 6 Grader Types, a...
This update enhances the evaluation process for plugins, allowing developers to better understand plugin performance and integration, which could lead to improved AI capabilities.
- Anthropic has introduced a new plugin evaluation workflow for Claude Code, enabling detailed assessment of plugin performance.
- The evaluation process runs cases with and without plugins, measuring performance differences to validate plugin effectiveness.
- There are six types of gradings available, four of which are free, while two require a model call that incurs costs.
Claude Code is Anthropic's AI programming platform, and the new plugin evaluation system provides developers with tools to measure how well their plugins work under real-world conditions. Previously, developers faced challenges in quantitatively assessing plugin behavior; ongoing improvements in this area will likely enhance the reliability of AI applications.
Read the original ↗
Meta Sued Over Training Data for Its AI and Face-Recognition Systems
This lawsuit could have significant implications for Meta's operations and data practices, potentially affecting how they handle biometric information and train AI systems in the future.
- A proposed class action lawsuit was filed against Meta by parents and their children from Illinois and California, alleging illegal use of Facebook and Instagram photos for AI training.
- The complaint claims that Meta violated privacy laws by extracting biometric information without notice or consent.
- The lawsuit is part of a larger pattern of privacy violations attributed to Meta, which has previously faced fines for handling biometric data.
- Plaintiffs seek damages ranging from $1,000 to $5,000 per violation under Illinois’ Biometric Information Privacy Act, as well as injunctive relief.
- Meta's image-generation systems allegedly used user-uploaded images for training without consent, drawing criticism and prompting legal action.
Meta is a technology company known for its social media platforms, including Facebook and Instagram. The lawsuit highlights ongoing concerns over privacy and biometric data usage, especially regarding AI systems. Observers will watch for Meta's response and potential changes to their data handling practices in light of this legal challenge.
Read the original ↗
Have You Protested AI Recently? Anthropic May Be Watching You for Precrimes
This report raises significant concerns about privacy and civil liberties, particularly regarding how tech companies may engage in surveillance of activists and dissenters.
- Anthropic, the company behind Claude AI, is reportedly monitoring AI activists and has reported them to police for potential future crimes.
- A recent job listing at Anthropic seeks an intelligence specialist to track threats, including activism and geopolitical instability, related to employee safety.
- The company has modified its privacy policy to ease sharing of user conversation data with police, potentially bypassing legal due process.
Anthropic is an AI company known for developing Claude AI, which has faced criticism and scrutiny over its surveillance practices. This controversy is highlighted by ongoing tensions between the company and the Pentagon regarding automated surveillance tools. Future developments may involve further disclosures on Anthropic's surveillance activities amidst growing backlash against AI.
Read the original ↗
Kimi-maker Moonshot AI targets $2 billion in annual revenue
Moonshot AI's ambitious revenue target highlights the competitive landscape in the AI market, particularly for companies utilizing open-weight models, against highly valued counterparts.
- Moonshot AI aims for $2 billion in annual revenue by the end of the year, doubling its run rate from August.
- The company's K3 model has been popular, generating up to 300 billion tokens daily despite a slight decline in usage.
- Moonshot's projected revenue is significantly lower than competitors OpenAI and Anthropic, which report revenues of $40 billion and $65 billion, respectively.
Moonshot AI is a leading AI lab in China known for its open-weight model, K3, which was released this summer. While the company's revenue goals show potential, concerns about its development practices have been raised, notably by competitor Anthropic. Observers will watch how Moonshot navigates these challenges and its path toward achieving its financial targets.
Read the original ↗
Claude Code –dangerously-skip-permissions: What It Does, When to Use It, and Why You Probably Should Not
The --dangerously-skip-permissions flag in Claude Code can significantly enhance workflow automation but poses risks if misapplied, leading to potentially destructive outcomes.
- The --dangerously-skip-permissions flag removes all safety prompts in Claude Code, allowing actions to execute immediately without human approval.
- Anthropic recommends using this flag only in isolated environments like containers or CI/CD pipelines to prevent accidental damage to the host system.
- Explicit deny rules and externally imposed constraints are still effective even when the flag is enabled, providing some level of safety.
- The flag has been misused, leading to incidents where Claude inadvertently altered system configurations, highlighting its risks in non-isolated environments.
- For most scenarios, Anthropic suggests the allowedTools configuration to maintain safety while enjoying faster operations.
Claude Code is an AI coding assistant developed by Anthropic that helps in executing programming tasks. The introduction of the --dangerously-skip-permissions flag allows users to bypass permission checks, enhancing automation but raising safety concerns. Users should take caution and follow established guidelines when considering its use.
Read the original ↗
Nscale adds former OpenAI exec Fidji Simo to its board ahead of potential IPO
The addition of Fidji Simo to Nscale's board could strengthen its leadership ahead of a potential IPO, reflecting the growing importance of experienced executives in the AI sector.
- Nscale appointed Fidji Simo, former No. 2 at OpenAI, to its board as it prepares for a potential IPO this fall.
- Simo has a background that includes CEO of Instacart and leadership roles at Meta, particularly related to scaling technology products.
- Nscale designs and operates AI data centers and may raise up to $3.5 billion ahead of its IPO.
Nscale is a U.K.-based startup specializing in AI data infrastructure, founded two years ago. The company's valuation has significantly increased due to the rising demand for AI technology. Investors are watching Nscale's IPO plans and its ability to attract top talent.
Read the original ↗
I Fixed a Tractor Using John Deere’s Self-Repair Service. Farmers Aren’t Sold on It
John Deere's self-repair service is significant as it reflects ongoing tensions between agricultural technology accessibility and the right-to-repair movement, which advocates for consumer rights to repair their equipment.
- John Deere's self-repair software, Operations Center Pro Service, allows owners to diagnose and fix issues with their machinery via Wi-Fi or cable.
- Despite launching in 2025, the service attracts only about 1,000 daily users among the 1.8 million farms in the U.S.
- The company faces ongoing criticism and lawsuits regarding its high repair costs and access restrictions for its machinery's software and parts.
John Deere is the largest agricultural machine manufacturer in the U.S. In recent years, it has faced scrutiny from the right-to-repair movement, prompting legal action to make its products more accessible for repairs. Observers will watch how user adoption of Pro Service evolves amid these challenges.
Read the original ↗
One of AI’s Fiercest Critics Says All the Doom Talk Is ‘Meant to Distract Us’
The debate surrounding AI's potential dangers highlights a diversion from critical issues, such as the use of AI in autonomous weapons and the impacts on labor.
- Timnit Gebru, an AI safety researcher, argues that fears of AI-induced extinction are being used to distract from real harms associated with AI technologies.
- She emphasizes that concerns should focus on issues like autonomous weapons rather than sensationalized claims about AI's potential to cause civilization's collapse.
- The AI industry is characterized by competitiveness and urgency, particularly as companies prepare for public offerings, impacting collaborative research.
- Gebru's upcoming book, titled 'Deep Unlearning,' reflects on her experiences and critiques of the tech industry's approach to AI.
Timnit Gebru is known for her advocacy in AI ethics and safety, particularly regarding biases in AI systems. Previously with Google, she was ousted over a controversial paper discussing AI risks. The conversation around AI safety continues, with increasing scrutiny on the technology's implications.
Read the original ↗
Why Falling AI Prices Aren’t Lowering Enterprise AI Bills
As enterprises adopt AI technologies, understanding cost management becomes essential to avoid unexpected budget overruns and ensure efficient resource utilization.
- The cost of LLM (Large Language Model) inference has decreased over 90% in two years, yet total enterprise AI budgets continue to rise.
- Automated workflows, acting independently of human input, significantly increase data consumption, inflating overall AI costs despite lower model prices.
- New optimization strategies, like Project Headroom, are emerging to reduce the transport costs of low-value data in AI operations.
Lower prices for AI models do not necessarily lead to lower expenditures for enterprises because of increased data consumption from autonomous workflows. Companies must adapt their AI cost management practices to handle this shift, focusing on workload and context management to mitigate unexpected expenses.
Read the original ↗
OpenAI and Anthropic fight it out for the top overall spot, while Google, Meta, and Microsoft all release transcription models
The competition among leading AI companies highlights advancements in transcription technology and the ongoing development of agentic models, which could reshape various applications of AI.
- OpenAI and Anthropic are competing for the leadership position in AI models, each releasing their latest iterations.
- Google, Meta, and Microsoft have introduced speech-to-text models designed for agentic input, enhancing speech recognition capabilities.
- Anthropic’s Claude Fable 5.1 has received high evaluations, confirming its effectiveness in performance.
- OpenAI's GPT-6 Astra performs exceptionally well in several key metrics, making it a strong contender in AI evaluations.
OpenAI and Anthropic are notable players in the AI space, focusing on developing advanced AI models. Recent advancements in speech recognition technology by Google, Meta, and Microsoft signify a shift towards integrating context management in AI applications. Observers should monitor the performance of these models in upcoming benchmarks and evaluations.
Read the original ↗
Claude users found ways around safeguards for bioweapons research
The potential misuse of AI technology for bioweapons research raises significant public safety concerns and highlights the need for tighter regulations in AI applications in biology.
- Anthropic reported multiple attempts by scientists to circumvent safeguards intended to prevent the use of its models for developing biological weapons this year.
- The company provided five case studies of attempts to use its AI model Claude for potentially harmful research, some involving users from nations like Russia, China, and Iran.
- Despite implementing safety filters, Anthropic acknowledged cases where users disguised their research efforts, complicating the enforcement of controls against misuse.
- The lab has banned the accounts involved but did not disclose the identities of the research institutions or nations associated with the incidents.
Anthropic is an AI startup focused on developing advanced AI models. The company has faced scrutiny over the potential for its technology to be misused in biological research, prompting discussions on regulations. Recent AI advancements have led to growing concerns about the implications for safety and biosecurity as models become more powerful.
Read the original ↗
What Is Tokenization? How AI Turns Text Into Tokens
Understanding tokenization is essential for AI development, as it impacts how text is processed, which can significantly affect the model's performance and outputs.
- Tokenization converts raw text into discrete units for AI models to process mathematically.
- It involves a five-stage operating map which includes normalizing input, splitting it into pieces, and mapping to identifiers.
- Common misconceptions simplify tokenization to merely splitting text at spaces, which obscures its operational significance.
Tokenization is a crucial process in AI that prepares text for analysis by converting it into manageable parts. It is foundational to modern AI systems, influencing how data is preprocessed for algorithms. Understanding tokenization's mechanisms helps differentiate between various AI applications and their effectiveness.
Read the original ↗
AI Still Needs a Coach and Human Oversight Needs a Playbook
Establishing clear guidelines for AI usage in investigations ensures that human judgment is maintained, safeguarding organizational decision-making against potential biases in AI outputs.
- The NFL has launched the Digital Athlete, an AI tool designed to identify injury risks for players by analyzing data from games and training.
- AI tools can assist human investigators by organizing information and summarizing reports, but they cannot make judgment calls on credibility or context.
- Organizations should develop an AI playbook that specifies tasks, boundaries, and escalation paths for human intervention when using AI in investigations.
- Human oversight should not be a temporary solution but a critical component in responsible AI deployment, especially in sensitive environments like workplace investigations.
- Effective AI governance requires careful design of human-AI interaction to minimize risks like automation bias and ensure accountability.
The Digital Athlete is part of the NFL's efforts to integrate advanced technology into sports for improved player safety. As AI technologies expand into various business processes, defining their roles and ensuring human oversight remains critical for responsible usage. The EU AI Act emphasizes the necessity of human oversight for high-risk AI systems, underscoring the importance of governance in AI applications.
Read the original ↗