Bilingual

Breaking: Trump asks the impossible of Anthropic
13 hours ago
- The politics and inadequacy of guardrails for AI has become a central issue, as warned in January 2024.
- Security experts are correct about LLMs, which are fundamentally unsafe as next-token predictors.
- Guardrails for LLMs face a dilemma of being too restrictive or too permissive, and none have worked effectively.
- The problem is not specific to Anthropic but is a general issue with Generative AI.
- LLMs lack a moral compass, as they are based on matrix multiplication and can be manipulated to say anything given the right context.
- LLMs are a fatally flawed technology, unlike banks that can be saved with money, as we don't understand how they operate and their opacity increases with more resources.
- The failure of LLMs is logical and epistemological, as language is fundamentally different from thinking.
The Generative AI Fizzle™
13 hours ago
- The author warns about a potential 'Generative AI Fizzle' where AI stocks are overvalued and may decline slowly as hype outpaces actual profits.
- LLMs have become commodities with no technical moat, leading to price wars and elusive profitability for providers.
- A new open-source Chinese model could worsen the situation for US LLM companies.
- Debt is growing, and the bubble may already be fizzling, with significant month-over-month stock declines.
- Hardware sellers like Micron may benefit from the investment bubble but ultimately make losses worse for others when end-user demand fails to materialize.
- The author compares the mania to tulip mania, noting that the technology may persist but astronomical valuations will not.
Saga: Source Attribution of Generative AI Videos (identifies the model used)
3 days ago
- SAGA is the first comprehensive framework for large-scale source attribution of AI-generated videos, identifying the specific generative model used rather than just detecting real vs. fake.
- It provides multi-granular attribution across five levels: authenticity, generation task (e.g., T2V/I2V), model version, development team, and precise generator.
- The framework uses a novel video transformer architecture built on a robust vision foundation model to capture spatio-temporal artifacts.
- A data-efficient pretrain-and-attribute strategy allows SAGA to achieve state-of-the-art attribution using only 0.5% of source-labeled data per class, matching fully supervised performance.
- Temporal Attention Signatures (T-Sigs) offer a novel interpretability method that visualizes learned temporal differences, explaining why different video generators are distinguishable.
- Extensive experiments on public datasets, including cross-domain scenarios, demonstrate that SAGA sets a new benchmark for synthetic video provenance, providing interpretable insights for forensic and regulatory applications.
Gatekeepers, Luddites, Haters and AI-Generated Maps
3 days ago
- The author opposes generative AI in cartography, rejecting accusations of being a gatekeeper, Luddite, or AI hater.
- AI-generated maps are promoted by professionals, but the author, a non-mapmaker, refuses to use them because the AI does the work, not the user.
- Key issues: AI-generated maps lack copyright protection, contain serious inaccuracies (e.g., maps with two Ohios), and are shared without disclaimers, leading to ethical concerns.
- AI adoption in other fields has led to mass layoffs and precarious work, not leisure; professionals should be Luddites opposing poor conditions.
- Comparing generative AI to typewriters is flawed; the author likens it to cocaine—addictive and impairing creative thinking.
- Broader costs include environmental damage, plagiarism lawsuits, bubble economics, and public fury over AI replacing human interaction.
The AI Productivity Illusion
3 days ago
- Generative AI hasn't improved productivity at the company or economy level despite enabling faster task completion, as true economic productivity depends on output quality and input costs, not just speed.
- The AI infrastructure build-out already costs $1.5 trillion, and generative AI's error-proneness—like fabrications or errors in code/medicine—reduces output value and increases quality control costs.
- AI's economic value is limited because demand for its outputs isn't as high as transformative technologies like the Industrial Revolution, and it may harm productivity in fields like cybersecurity by enabling attacks.
- The apparent 'productivity paradox' may be an illusion: generative AI's gains from speed are offset by value destruction from unreliability, leading to no net economic improvement.
Codeberg Divides
4 days ago
- Codeberg updated its terms to exclude projects mostly written with generative AI, which the author views as within its rights but not necessarily wise or inclusive.
- Democracy in infrastructure does not guarantee predictability or neutrality; a democratic provider without a clear constitution may be less reliable than a corporation.
- The term 'mostly' in the policy is ambiguous and hard to enforce, making it difficult to determine authorship percentages in active codebases.
- A clearer rule—either banning all LLM involvement or addressing spam and resource abuse directly—would be preferable to the current middle ground.
- The open source community is deeply divided over LLMs and agents, yet these tools are becoming integral to software development and need constructive engagement.
- The author wishes Codeberg were more forward-looking, aiming to host future open source software rather than only software made in currently approved ways.
Computer science enrollment fell for the first time in 20 years
4 days ago
- Computer science enrollment declined for the first time in roughly two decades in the 2025-26 academic year.
- Enrollment in Computer and Information Sciences at four-year colleges fell 8.1% year over year in fall 2025.
- The decline coincides with the rapid adoption of generative AI like ChatGPT, but the data does not prove AI caused it.
- Stanford economist Jacob Light's research shows a 4.6% year-over-year drop in CS enrollment at the average school, with share falling 6.3 percentage points.
- Five possible explanations include a weaker job market, uncertainty about CS degree value in the AI era, weaker academic preparation, demographic changes, and expansion of competing fields.
- Despite the decline, enrollment remains above 2022 levels, suggesting the drop follows years of rapid growth.
Three Years of AI on Steam
13 days ago
- Valve's AI disclosure mandate started in January 2024, and by mid-2026, about one in three new Steam games included an AI disclosure.
- Monthly Steam releases grew significantly, with 60–90% of the increase from 2023 to 2026 attributed to AI-disclosed games, while non-AI launches rose only modestly.
- AI-disclosed games account for a smaller share of sales (10–27% in late 2025–2026) than their share of launches (~33%), indicating that more AI games are being released, but they are not individually more successful.
- Most AI games are commercial failures, with the median game earning around $300; AI revenue is highly volatile and dependent on a few hits.
- Disclosed AI use is predominantly visual, especially among unsuccessful games, while successful games are more likely to use AI for voice, localization, and text, and they provide more detailed and reassuring disclosures.
- About 22% of significant-revenue AI games added the disclosure flag retroactively, often around Valve's policy update in January 2026, indicating compliance-driven relabeling.
Scientists' Side Hustle? Using AI and Quantum Computing to Generate New Peptides
16 days ago
- Researchers demonstrated that quantum computers can enhance the accuracy and scope of generative AI models for drug discovery.
- A hybrid quantum-classical system accelerated AI to generate novel peptides, improving binding prediction to target proteins, especially with scarce training data.
- The project was self-funded by the team using spare time and leftover money, as innovative science is often deemed risky by traditional funders.
- The approach could speed up personalized immunotherapies, vaccines, and drug efficacy for understudied populations, addressing data gaps in non-Western genetic diversity.
- Quantum computing remains nascent and faces skepticism; current limitations include small scale and inability to run full AI models, but the study offers a near-term commercial application example.
- Future directions include applying the workflow to more advanced models and larger proteins, and exploring uses like designing synthetic antidotes for snakebite venom.

Hasty Briefsbeta

generative ai

Breaking: Trump asks the impossible of Anthropic

The Generative AI Fizzle™

Saga: Source Attribution of Generative AI Videos (identifies the model used)

Gatekeepers, Luddites, Haters and AI-Generated Maps

The AI Productivity Illusion

Codeberg Divides

Computer science enrollment fell for the first time in 20 years

Three Years of AI on Steam

Scientists' Side Hustle? Using AI and Quantum Computing to Generate New Peptides