Hasty Briefsbeta

Bilingual

'Incredible' milestone reached as Sweden becomes a smoke-free country
2 hours ago
- Sweden achieved its political goal of becoming smoke-free by 2025, with regular smokers dropping to below 5%.
- The proportion of daily smokers decreased from 16% in 2003 to 4.8% in 2025, and people are smoking fewer cigarettes overall.
- Decline in smoking attributed to reduced accessibility through higher taxes and advertising bans, making it more expensive and harder to smoke.
- Snus use increased in Sweden, but research indicates it may not be the primary factor for reduced smoking rates.
- Preventative measures are credited as the main driver behind the reduction in smoking habits.
- Experts note a need for more research into snus's role in smoking trends.
The Ballad of TIGIT
2 hours ago
- Amyloid-beta drugs for Alzheimer's are widely seen as disappointments, though their story is nuanced, with potential limited early-stage benefits.
- TIGIT drugs emerged from cancer immunotherapy research, inspired by Keytruda's success, as a promising new immune-system brake target.
- TIGIT inhibitors were theorized to block immune checkpoints more effectively than Keytruda, leading to massive investments and parallel trials by companies like Roche and Merck.
- Initial phase 2 results for Roche's tiragolumab showed promise, but subsequent phase 3 trials failed, including a flagship lung cancer trial missing progression-free and overall survival.
- Merck's vibostolimab program also failed due to safety issues and lack of efficacy, leading to discontinuation.
- An Fc-silent antibody hypothesis (to avoid killing cancer-fighting T-cells) was tested with domvanalimab but also failed in phase 3.
- Other companies like BeiGene, GSK, and Novartis faced similar failures, with billions spent and thousands of patients enrolled in trials.
- AstraZeneca's rilvegostomig remains as a last hope, but prospects are dim given prior failures.
- The broader failure is attributed to biological complexity: TIGIT's role in immune signaling may be too context-dependent for a one-size-fits-all therapy.
- Despite solid scientific rationale and early signals, TIGIT drugs exemplify how even well-founded drug candidates can fail unexpectedly in clinical development.
Language Models Need Sleep
2 hours ago
- Transformers' attention struggles with long contexts due to scaling issues.
- A sleep-like mechanism consolidates recent context into fast weights via offline recurrent passes.
- Sleep allows extra computation without affecting real-time inference latency.
- Tested on synthetic and math reasoning tasks where standard models fail.
- Performance improves with longer sleep, especially for deeper reasoning requirements.
Stockholm poised to become leading European geospatial intel player
3 hours ago
- French intelligence industry joined Macron's delegation to South Korea and Japan.
- Efforts to reduce ties with the US opened opportunities.
- Very few solid contracts materialized during the visits.
Launch HN: Minicor (YC P26) – Windows desktop automations at scale
2 hours ago
- Minicor provides a platform for building desktop automations using self-healing agents.
- Agents adapt to UI changes and verify actions to maintain automation reliability.
- RPAs can be deployed via API call on Windows VMs, cloud, or on-premise.
- Built-in observability includes video recordings, Slack notifications, and debugging tools.
- Minicor addresses automation in legacy systems lacking APIs, common in healthcare, automotive, and finance.
- It offers faster deployment (weeks vs. months) and higher accuracy (93-96%) than traditional RPA.
- The platform is SOC 2 Type II certified and HIPAA compliant, with on-premise deployment options.
Don't Subscribe So Casually
3 hours ago
- Subscriptions act like recurring votes on identity, similar to roommates, shaping who you become over time.
- Chatbots and AI subscriptions amplify addictive patterns and negative effects, making conscious evaluation crucial.
- Unlike one-time goods, subscriptions are future-oriented, expanding options but also altering future behavior substantially.
- Even worthwhile subscriptions, like Uber One or Costco, shift consumer preferences and behaviors subtly.
- Subscriptions create structural disadvantages for consumers, as companies optimize for long-term profitability and behavioral modification.
- Choosing subscriptions should align with company values, as their goals directly influence user experience and outcomes.
- AI subscriptions are particularly insidious due to manipulative interactions and unclear long-term business models.
- Without active self-determination, corporate incentives can shape personal choices and psychological makeup.
About That Erdős Problem
6 hours ago
- OpenAI's announcement of a counterexample to the Erdős unit distance conjecture in under five pages, with comments from nine prominent mathematicians, sparked discussion.
- The week featured significant AI-related news, including OpenAI and Anthropic gearing up for IPOs, SpaceX's filings, and Eric Schmidt facing backlash at college commencements.
- Concerns about AI's societal impact were highlighted, such as potential political backlash by 2028, regulatory delays, and issues like bad code generation and environmental costs.
- Mathematicians' responses varied: Victor Wang questioned the social contract between mathematicians and AI companies regarding freely available work being used for financial gain.
- Melanie Matchett Wood urged caution about drawing conclusions from AI announcements without context on AI errors, while Tim Gowers noted efficiency and environmental concerns.
- The author emphasized the need for mathematicians to engage in public debate on AI, referencing quotes on ownership, control, and profit-sharing from AI built on collective data and work.
- The piece calls for building a social contract between mathematics and AI, viewing AI as a source of new challenges and connections to benefit mathematical research and ethics.
Opaque Types in Python
3 days ago
- Opaque data types in Python using typing.NewType help manage complex configuration objects with minimal public API exposure.
- They allow private internal representation changes without breaking client code, preserving forward compatibility.
- Public constructor functions (e.g., shipFast) provide a stable interface while hiding implementation details.
- This pattern is useful for libraries where configuration state is complex and likely to evolve, like shipping options with multiple carriers and methods.
The Open/Closed Problem in AI
10 hours ago
- The author attended the MLSys conference and observed a focus on improving efficiency in training and deploying LLMs.
- The Open/Closed problem is illustrated through the evolution of computer graphics: from open CPUs to closed fixed-pipeline GPUs, back to open programmable GPUs, and now to closed specialized ASICs for AI inference and training.
- Current AI systems use open-loop learning, where models are trained externally and don't learn after deployment, unlike the closed-loop learning of the human brain, which updates itself internally based on sensory predictions.
- The AI industry is heavily investing in optimizing open-loop learning through hardware like ASICs and kernel improvements, which inadvertently makes closed-loop learning more difficult to achieve.
- Specialized hardware for inference assumes static models, with read-only parameters and separated compute and memory, contrasting with the needs of closed-loop learning, which requires dynamic weights, fine-grained updates, and fused memory-compute.
- The author warns that focusing on open-loop efficiency may hinder breakthroughs in closed-loop learning, as hardware becomes increasingly specialized around the current paradigm, limiting experimentation and innovation.
Incident with Actions and Pages
6 hours ago
- Incident resolved after addressing authentication and performance issues affecting GitHub Actions and Pages.
- Root cause analysis to be shared upon availability.
- Degradation mitigated and stability being monitored.
- Incident investigation initiated due to reported performance degradation.
Netherlands blocks US takeover of vital digital supplier
6 hours ago
- Dutch government blocks U.S. firm Kyndryl from acquiring Dutch IT supplier Solvinity.
- Solvinity operates the DigiD app, crucial for citizens' online authentication in the Netherlands.
- Acquisition raised concerns about foreign control over key digital infrastructure and Europe's reliance on U.S. technology.
AWS Fired the One Employee Who Gave a Damn
4 hours ago
- An AWS employee who helped restore a deleted account was fired despite his efforts.
- The employee's internal escalation led to a CEO-level investigation, but systemic issues remained unfixed.
- AWS's shift toward AI and automation is undermining human-driven customer support and accountability.
- Layoffs and terminations at AWS have led to a talent drain, with many former employees leaving tech entirely.
- There is a pattern of AWS blaming human error for incidents that involve automated systems or AI.
- The 'Correction of Error' process launched after the account deletion failed to address key issues.
- The article criticizes big tech for valuing efficiency over human empathy and genuine problem-solving.
- The open-source community's credibility is harmed when companies fire employees who earned trust.
- The author urges readers to maintain backups and exit strategies from cloud providers.
- The overall message is about the erosion of humanity in technology and corporate systems.
US Law Enforcement Warns of 'Anti-Tech Extremism'
6 hours ago
- Federal intelligence agencies and law enforcement are shifting focus to monitor 'anti-technology extremists' as an emerging domestic threat.
- Over 1,000 pages of reports from DHS, FBI, and fusion centers reveal surveillance on activities and beliefs challenging AI and data center proliferation.
- A New York Intelligence Bureau report warns of potential civil unrest due to AI adoption, coining the term 'anti-tech violent extremism.'
- Fusion centers are tracking protests and dissent related to technology, including events like 'Tesla Takedown' and data center opposition groups.
- Private intelligence firms like SITE are monitoring online sentiment, sometimes flagging nonviolent content as suspicious or threatening.
- Legal experts criticize vague suspicious activity reports, arguing they can criminalize peaceful protest and free speech.
- The framework of 'anti-tech extremism' risks encompassing peaceful activists, AI skeptics, and critics of technology's societal impact.
A reality check on the AI jobs hysteria
6 hours ago
- Despite fears of AI causing mass job losses, current economic research shows little evidence of large-scale impact on the US labor market so far.
- Unemployment rates for jobs most exposed to AI are lower than those in less exposed occupations, and no significant shifts to safer jobs are observed.
- Only about 20% of companies use AI in business functions, indicating slow adoption and suggesting disruption is not immediate, allowing time for planning.
- Young workers (ages 22-25) in AI-exposed fields like software development face job declines, but this represents a small portion of the overall labor market.
- AI may be automating entry-level tasks (codified knowledge), while augmenting roles requiring experience (tacit knowledge), potentially breaking the 'earn-while-you-learn' model.
- Coding jobs continue to grow but at a slower pace, with wages in AI-exposed sectors rising, implying demand for skills hard to replace with AI persists.
- Predictions of AI-driven job apocalypses (e.g., radiologists, driverless trucks) have historically been inaccurate, overlooking job complexity and adaptation.
- Better data is needed to understand AI's impact, as current tools fail to capture granular effects on occupations, skills, and workplace transitions.
- The speed of AI disruption is critical: gradual change allows labor market adaptation, while sudden shifts could challenge policymakers and workers.
- Investment in understanding AI's societal impacts lags behind spending on the technology itself, hindering preparedness for economic transformations.
Exposing Critical Vulnerabilities in CBSE's On-Screen Marking Portal
8 hours ago
- The author discovered critical vulnerabilities in CBSE's On-Screen Marking (OSM) portal, used for Class 12 board exam evaluations.
- Vulnerability 1: A hardcoded master password in the frontend JavaScript bundle allowed bypassing OTP and logging in as any examiner.
- Vulnerability 2: OTP validation was performed client-side, making it trivial to extract or bypass.
- Vulnerability 3: No route guards in the Angular app enabled direct navigation to internal pages without authentication.
- Vulnerability 4: The password change feature did not require the old password, allowing reset of any examiner's password.
- Vulnerability 5: Systemic IDOR across the API let attackers act as any user by manipulating browser storage values.
- These flaws combined enabled full account takeover, mark tampering, and disruption of the grading process.
- The vulnerabilities were reported to CERT-In but remained unpatched for a long time, prompting public disclosure.
- Key lesson: Client-side code cannot be trusted; secrets and security decisions must be enforced server-side.
Don't put aria-label on generic elements like divs
4 days ago
- Avoid placing aria-label on generic elements like divs or spans.
- The ARIA spec prohibits naming generic roles; labeling them leads to inconsistent screen reader announcements.
- Screen readers vary: some announce the label, others ignore it, and some announce content.
- Exceptions include section elements (role changes to region) and popovers (role changes to group).
Eagle 3.1: Collaboration Between the EAGLE Team, vLLM Team, and TorchSpec Team
6 hours ago
- EAGLE 3.1 is introduced by the EAGLE team, vLLM team, and TorchSpec team, advancing speculative decoding.
- It addresses performance degradation issues like attention drift in various deployment scenarios.
- Key architectural improvements include FC normalization and feeding post-norm hidden states into next decoding steps.
- EAGLE 3.1 shows better extrapolation, long-context robustness, resilience to variations, and stable acceptance lengths.
- In long-context workloads, it achieves up to 2× longer acceptance length compared to EAGLE 3.
- TorchSpec provides efficient training support for EAGLE 3.1 and future speculative decoding algorithms.
- An open-sourced EAGLE 3.1 draft model for Kimi K2.6 is released, demonstrating deployment with TorchSpec and vLLM.
- vLLM integrates EAGLE 3.1 with features like FC normalization and backward compatibility with existing checkpoints.
- Benchmarking shows significant throughput improvements, such as 2.03× higher output throughput at low concurrency.
- The collaboration highlights open-source efforts across algorithm research, system optimization, and training infrastructure.
GitHub Actions down again today
6 hours ago
- Service delays occurred from 13:41 to 17:43 UTC on May 12, 2026, affecting Code Scanning, notifications, webhooks, and Slack integrations.
- 53% of Code Scanning check runs exceeded 15 minutes; notifications averaged 22 minutes, and Slack webhooks averaged 20 minutes in delivery.
- The primary cause was replication lag from an internal database migration, leading to insufficient worker capacity for high job enqueues.
- Mitigation involved scaling processing workers to handle increased load, restoring normal service times.
- Future prevention plans include creating dedicated worker pools for high-usage shared queues.
- Recovery updates show services, including Webhooks and CodeQL, returning to normal by 17:43 UTC.
China vs. Taiwan: The Geography of an Unfinished War
7 hours ago
- Taiwan is a strategic hinge connecting the East and South China Seas, controlling major sea lanes that link Northeast Asia to Southeast Asia and beyond.
- For China, Taiwan represents unfinished civil war business and a barrier along the first island chain limiting its naval access to the wider Pacific.
- Taiwan's dependence on imported energy (LNG and coal) makes maritime security critical, with disruptions potentially causing economic and social crises.
- Taiwan is central to the global semiconductor ecosystem, making advanced chips instruments of industrial and military power.
- The conflict embodies a clash between China's authoritarian model, Taiwan's democratic model, and the U.S.-led alliance-based Indo-Pacific order.
- If China absorbed Taiwan, it would alter Asia's military geography, exposing Japan's flank and increasing pressure on the Philippines.
- Future scenarios include managed tension, coercive strangulation via cyber and economic pressure, or open military conflict through blockade or invasion.
- Taiwan's fate hinges on ports, energy reserves, naval logistics, chip supply chains, alliance credibility, and democratic resilience under pressure.
- The dispute is a structural contest between continental control and maritime openness, ideological absorption versus democratic continuity.
The AI bubble isn't like the internet bubble
7 hours ago
- The early web saw workers eagerly adopting new tools like Lotus Notes, which acted as a compromise between IT control and user flexibility, contrasting with today's AI where bosses force usage on reluctant workers.
- Workers historically smuggled in technologies like Visicalc and web-based tools to enhance productivity, while AI adoption is driven top-down, with companies monitoring and enforcing its use.
- AI differs fundamentally from the early web in unit economics; the web grew more profitable with each user, while AI becomes less profitable, described as having 'dogshit unit economics.'
- Teenagers and young adults, typically early adopters, show aversion to AI, unlike the web era where they were passionate advocates demanding workplace integration.
- Some autonomous workers adopt AI voluntarily to improve quality (as centaurs), but many are coerced into using it as 'reverse-centaurs,' serving machines rather than being aided by them.

first prev0next

About|Login

Hasty Briefsbeta

'Incredible' milestone reached as Sweden becomes a smoke-free country

The Ballad of TIGIT

Language Models Need Sleep

Stockholm poised to become leading European geospatial intel player

Launch HN: Minicor (YC P26) – Windows desktop automations at scale

Don't Subscribe So Casually

About That Erdős Problem

Opaque Types in Python

The Open/Closed Problem in AI

Incident with Actions and Pages

Netherlands blocks US takeover of vital digital supplier

AWS Fired the One Employee Who Gave a Damn

US Law Enforcement Warns of 'Anti-Tech Extremism'

A reality check on the AI jobs hysteria

Exposing Critical Vulnerabilities in CBSE's On-Screen Marking Portal

Don't put aria-label on generic elements like divs

Eagle 3.1: Collaboration Between the EAGLE Team, vLLM Team, and TorchSpec Team

GitHub Actions down again today

China vs. Taiwan: The Geography of an Unfinished War

The AI bubble isn't like the internet bubble