Hasty Briefsbeta

Bilingual

Musk demanded proof people died from USAID cuts. He got it – and lost it
an hour ago
- Elon Musk challenged critics to name people who died due to USAID funding cuts he led, claiming no deaths occurred.
- Journalist Nicholas Kristof provided names of individuals, including children like Jibia and Achol Deng, who died after losing USAID-funded healthcare.
- Musk responded with insults, calling Kristof 'evil' and a 'liar,' and targeted others like researcher Alonso Gurmendi.
- Kristof published a New York Times column detailing deaths linked to aid cuts, with estimates suggesting over 750,000 lives lost globally.
- Musk bragged about cutting USAID funding and faced legal issues, including fraud lawsuits related to a $1 million giveaway.
- Meta faced backlash for using Kylie Jenner in smart glasses ads, highlighting wealth disparity and privacy concerns over surreptitious recording.
- Meta contractors posed as children to test rival AI chatbots on sensitive topics like suicide and eating disorders, raising ethical questions.
- Mark Zuckerberg plans high-quality beef production on his ranch and commented on AI job displacement, despite recent Meta layoffs.
- Larry Ellison donated $45 million to pro-Trump groups, with Trump supporting his media acquisition efforts, including potential CNN overhaul.
- Global billionaire wealth increased by 25%, while median wealth declined for the working class, according to a UBS report.
Say No to Online Age Verification
8 hours ago
- Age verification systems require submitting sensitive personal data to third-party companies to access certain online content, starting with adult sites and expanding to social media and beyond.
- Anonymous speech is essential for protecting vulnerable groups like whistleblowers and political dissidents; age verification leads to self-censorship and chills free expression.
- Privacy is compromised as data breaches expose IDs, selfies, and other personal information, with systems enabling surveillance under the guise of child safety.
- Laws mandating age verification have spread globally from Louisiana in 2023 to multiple countries, targeting increasingly broad areas like AI chatbots and VPNs.
- Children's rights to access information and free speech are undermined by age verification, which restricts their ability to explore ideas and seek knowledge online.
- Counterarguments highlight that age verification expands in scope, creates surveillance infrastructure, and fails to protect data, with alternatives like privacy laws and parental controls being more effective.
- Action steps include contacting lawmakers to oppose age verification mandates, using provided email templates to advocate for free expression and privacy.
Why implementing ActivityPub is hard, and why it doesn't have to be
8 hours ago
- Implementing ActivityPub from scratch is challenging due to complexities like handling multiple signature standards, JSON-LD shape variations, and ensuring secure defaults.
- Fedify is a TypeScript framework that abstracts away these complexities, providing built-in solutions for signatures, interoperability quirks, and delivery infrastructure.
- The framework supports various web frameworks and databases, includes development tools like CLI and debugging, and is used in production by platforms like Ghost and Hollo.
The SRE Guide to Kubernetes Observability: Red vs. Use Methods
7 hours ago
- The RED method, introduced by Tom Wilkie in 2015, focuses on request-serving services by measuring Rate, Errors, and Duration from the caller's perspective to assess service health.
- The USE method, defined by Brendan Gregg in 2012, applies to resources like CPU, memory, and disk, evaluating Utilization, Saturation, and Errors to identify resource bottlenecks.
- RED and USE address different layers: RED is service-oriented and blind to resource issues, while USE is resource-oriented and ignores service-level performance, highlighting their complementary roles in observability.
- Linkerd provides automatic RED metrics for meshed pods via its proxy, emitting response_total and response_latency_ms_bucket metrics without application instrumentation, covering east-west traffic.
- In an example incident, RED metrics showed rising latency while USE metrics revealed pod-level CPU throttling, demonstrating the need for both methods to diagnose issues comprehensively.
AI saves about 3% of your hours, and almost none of it reaches the money
3 hours ago
- AI saves about 2.8% of work hours in real-world settings, but these time savings rarely translate into higher earnings or profits.
- Task-specific AI gains are substantial (e.g., 40% faster writing, 14% more issues resolved per hour), but they shrink when applied across entire jobs due to tasks AI can't handle and verification overhead.
- AI's competence is 'jagged'; it performs well on structured tasks like writing and customer support but fails confidently on tasks outside its range, leading to costly errors.
- Only 3-7% of productivity gains from AI reach workers' pay, with most organizations seeing zero return on investment despite significant spending.
- To capture AI's value, focus it on high-volume, repeatable tasks it excels at, avoid using it outside its strengths, and deliberately convert saved time into revenue or cost savings.
I Wasn't Allowed Prompting ChatGPT During My Chalk Talk: This Is Discrimination
2 hours ago
- The author describes a chalk talk interview for a tenure-track academic position where they attempted to use ChatGPT to generate responses, which was met with disapproval from the committee.
- They argue that their reliance on AI tools like ChatGPT and Claude reflects modern scientific practice, where prompting and editing outputs are central to their workflow.
- The author criticizes the traditional chalk talk format as outdated, claiming it fails to assess real-world skills like AI-augmented cognition and prompt engineering.
- They defend their approach as collaborative, viewing AI as a co-investigator, and suggest that memorization is less relevant in contemporary research.
- After being rejected for concerns about independent thinking and foundational knowledge, the author is shifting to industry roles, which they perceive as more accepting of AI use.
- The piece is presented as satire, highlighting tensions between academic traditions and evolving technological practices in science.
Musk's X poses "serious risk to Americans' privacy," advocates warn FTC
6 hours ago
- Advocates urge the FTC to reject X's request to end audits and maintain oversight over its data handling.
- X argues that the FTC order is burdensome and unnecessary due to rebranding and overlapping GDPR obligations.
- Privacy advocates warn that X's current leadership poses risks to data security, necessitating continued FTC monitoring.
Trump has made more than $1B from crypto in a year
4 hours ago
- Donald Trump earned over $1 billion from cryptocurrency in the past year, despite previously calling it a 'scam'.
- Trump's crypto ventures include the $TRUMP memecoin, governance token WLFI, and stablecoin USD1, primarily operated through World Liberty Financial and affiliated companies.
- The $TRUMP memecoin generated over $600 million, while World Liberty brought in more than $500 million, contributing to Trump's estimated net worth rise from $2.3 billion in 2024 to $6 billion.
- USD1 stablecoin benefited from Binance's involvement and a $2 billion investment from Abu Dhabi's MGX, creating significant interest-earning reserves and regulatory leniency.
- Trump's crypto dealings have raised corruption concerns, with critics alleging conflicts of interest and untraceable gifts, while some regulations like the GENIUS Act are seen as positive but risk being undermined.
Instagram running ads promoting child sexual abuse material in India, BBC finds
3 hours ago
- Instagram ran paid ads promoting child sexual abuse material (CSAM) in India, linking to Telegram channels selling it for as little as 99 rupees.
- Meta's moderation system initially approved some ads, but later disabled several and suspended accounts after BBC investigation, while Telegram removed over 274,000 groups/channels in 2026.
- Indian authorities were notified, and critics accuse Meta of prioritizing revenue over safety, with former employees citing algorithmic promotion of extreme content and insufficient oversight.
- Meta denied knowingly targeting inappropriate ads and reported CSAM to the National Center for Missing and Exploited Children (NCMEC), but experts highlight gaps in cross-platform moderation and law enforcement challenges in India.
- Child sexual abuse in India is often linked to organized crime, with low reporting rates and police lacking technical skills; international cooperation is deemed crucial for tracking and combating such crimes.
BitTorrent Explained: Deep Dive into Peer-to-Peer Sharing, Security, and Usage
6 hours ago
- BitTorrent is a peer-to-peer protocol that decentralizes file sharing, boosting speed and resilience.
- It splits files into pieces distributed among users, with features like trackers, DHT, and encryption.
- While often associated with piracy, it has legal uses by organizations for distributing large files.
- Security risks include lack of anonymity and potential malware, with legal issues around copyrighted content.
- The protocol has evolved with BitTorrent v2, WebTorrent, and blockchain integration via BTT tokens.
Cache Merging as a Convergent Replicated State for Multi-Agent Latent Reasoning
6 hours ago
- Prior work (Agent Primitives) concatenates KV-caches for multi-agent latent reasoning with BagMerge, which is non-commutative and depends on unpredictable input ordering.
- CanonicalMerge makes the exchange a convergent replicated state by ordering caches by mean K-norm at a middle layer, ensuring byte-identical merged cache under any input permutation.
- The replicated state is separated from decode-time layout, using content-addressed latent fragments with set union as a state-based CvRDT, where CanonicalMerge is its deterministic render.
- CanonicalMerge matches the best BagMerge ordering across regimes and budgets without prior knowledge, trading minimal accuracy loss for structural guarantees.
- On multi-document QA (HotpotQA), CanonicalMerge outperforms output-fusion baseline PackLLM by 45 points at matched budget, showing cache-level merging as distinct from output-level fusion.
AI's $2.2T deficit fix is already half fake, economists say
3 hours ago
- AI could reduce the U.S. annual budget deficit by $2.2 trillion (from 6% to 2% of GDP) by 2036 through productivity gains.
- However, over half of these savings may be offset by five negative side effects: longer lives increasing entitlement costs, tax base shifts favoring lower-taxed capital gains, a weaker labor force reducing tax revenue, higher borrowing costs due to interest rates, and an AI arms race raising defense spending.
- AI-driven healthcare improvements could lower mortality but expand the retirement-age population, increasing Social Security and Medicare expenses.
- Productivity gains may accrue more to asset owners than workers, leading to a narrower tax take despite economic growth.
- AI could reduce labor force participation by 3% (equivalent to 6 million fewer workers by 2036), increasing reliance on government support programs.
- Increased investment in AI infrastructure might raise interest rates, adding around $60 billion to federal debt-service costs by 2036.
- An international AI arms race could add over $350 billion in cumulative defense spending over the next decade.
- Overall, net savings from AI might be closer to $1 trillion or less, emphasizing that AI is not a substitute for long-term fiscal reforms like tax hikes or entitlement cuts.
Heuristics for lab robotics, and where its future may go
6 hours ago
- There are two main types of lab robots: 'boxes' (like liquid handlers) that automate specific tasks and 'arms' (like the UR5) that move materials between boxes to form integrated workcells.
- Most lab protocols can be automated, but often aren't worth automating because the upfront investment in time and cost is only justified for high-throughput, repetitive tasks, not for exploratory research that changes frequently.
- Three approaches to improve lab automation are: better translation layers (e.g., Synthace, Briefly Bio) to convert human intent into robot instructions, better physical hardware integration (e.g., Automata, Ginkgo Bioworks) to connect systems, and adding intelligence (e.g., Medra, Zeon Systems) for error recovery and adaptability.
- The long-term business model for lab automation likely converges on 'cloud labs' (like the defunct Transcriptic or Emerald Cloud Labs), which centralize experiments to maximize robot utilization and can vertically integrate to reduce costs, though early attempts struggled by being too general and offloading logistics to users.
- The ultimate value of lab robotics may not be just increasing throughput, but enabling better, more predictive assays (improving 'predictive validity') and higher-standard experimentation at scale, potentially transforming drug discovery.
Costco Is the Anti-Amazon
5 hours ago
- Amazon and Costco represent two contrasting retail models: Amazon focuses on near-infinite assortment and ultra-fast home delivery, while Costco offers constrained selection and in-store shopping with low SKU counts.
- Costco's low SKU count enables faster inventory turnover, low overhead costs, higher wages for workers, low labor turnover, and a negative cash conversion cycle without pressuring suppliers.
- The logistical simplicity of Costco's model, including cross-docking pallets and minimal automation, is more efficient and socially beneficial than Amazon's complex delivery network, reducing infrastructure demands and environmental impact.
- New York City's public grocery initiative can learn from Costco by adopting low SKU counts, high-volume sales, central warehousing, supplier relationships, and loss leaders to keep prices low and overhead minimal.
- Historical context shows Costco's origins trace back to Fedco, a membership store for federal employees, highlighting the power of collective buying and public models for retail efficiency.
Markets are competitive if and only if P != NP
4 hours ago
- The paper argues that competitive markets depend on computational intractability: if P = NP, firms can efficiently detect collusion, making it sustainable; if P ≠ NP, detection is infeasible under certain conditions, preventing collusion stability.
- It links this to Maymin (2011), concluding that markets cannot be both informationally efficient and competitive, forming a fundamental trade-off.
- Artificial intelligence enhances firms' computational power, shifting markets toward collusion and explaining algorithmic collusion without explicit coordination.
International chess federation sanctions Kramnik
3 hours ago
- The FIDE Ethics & Disciplinary Commission (EDC) found former World Champion Vladimir Kramnik responsible for multiple violations of the FIDE Ethics and Disciplinary Codes.
- The violations included breaches related to dignity, respectful treatment, bullying, cyberbullying, psychological abuse, and failure to cooperate with a Fair Play Commission investigation.
- Several other charges, such as those concerning integrity, honesty, and reputational harm to FIDE, were dismissed due to insufficient evidence.
- The EDC emphasized that combating cheating is a top priority but must be handled through confidential procedures with proper evidence, not public accusations.
- Kramnik received a two-year ban from FIDE competitions, with the final 12 months suspended for a three-year probation, resulting in an active one-year suspension if no further breaches occur, plus 12 months of unpaid community service.
Evaluation order and nontermination in query languages
2 days ago
- λFS combines functional programming with relational programming, treating relations as finite functions.
- Nontermination in recursion introduces challenges with evaluation order and semantics.
- Three evaluation strategies discussed: left-to-right, nondeterministic, and parallel 'and'.
- Left-to-right evaluation offers predictability but may be inefficient for cyclic queries.
- Nondeterministic evaluation allows optimization but complicates compositional reasoning.
- Parallel 'and' restores determinism and symmetry but is difficult to implement efficiently.
Minimum Connection Times at Airports Worldwide
a day ago
- The article discusses Minimum Connection Times (MCT) for airports, offering tips and terminal maps for stress-free plane changes.
- It lists the top 10 busiest airports worldwide, including Hartsfield-Jackson Atlanta (ATL), Dubai International (DXB), and London Heathrow (LHR).
- For O'Hare International (ORD), Dallas/Fort Worth (DFW), and Denver International (DEN), specific MCTs are provided: 40 minutes for domestic to domestic, 90 minutes for other transfers.
- The article also includes a browse section listing countries with airports and their counts, such as Afghanistan (28) and Australia (584).
3D Printing a Brain (2021)
a day ago
- The project involved creating a 3D-printed brain trophy from an MRI scan in 2021.
- The process included converting DICOM to a mesh, cleaning in Blender, printing halves on an SLA printer, casting silicone molds, pouring pewter, and assembling with epoxy.
- A custom base was made using a router and chamfer bits, with a 3D-printed plaque.
- Key learnings included medical file types, new manufacturing techniques, and achieving a polished final result.
What we learned building a multi-agent PDF table extractor
5 hours ago
- Extraction demos are easy but production-level extraction must handle unseen documents, human-centric layouts, and high volumes with 99%+ accuracy.
- Real-world challenges include diverse layouts for the same document type, lengthy tables exceeding LLM token limits, and designs optimized for human reading.
- A single autonomous LLM agent is too costly and slow for production; instead, a multi-agent pipeline splits tasks among six specialized agents.
- The pipeline uses preprocessing to convert documents into per-page layout-preserved text and images, enabling page-by-page processing.
- Agent #0 generates and caches prompts for specific table classes, leveraging LLM knowledge for zero-shot generalization.
- Agent #1 acts as a cheap table presence detector, filtering out irrelevant pages to reduce costs by up to 80%.
- Agent #2 extracts table metadata, recovering complex structures like headers, subtables, and row extra data.
- Agent #3 extracts content with full provenance, using OCR text as the source of truth and images for spatial reference.
- Agent #4 generates Python code to map extracted data to the user's target schema, enabling deterministic transformations and synthetic columns.
- Agent #5 executes the generated code in a sandboxed environment, ensuring security and deterministic output without additional LLM calls.
- Key production-ready features include caching, provenance tracking, and pushing deterministic tasks out of LLMs into code for reliability and cost efficiency.

first prev0next

About|Login

Hasty Briefsbeta

Musk demanded proof people died from USAID cuts. He got it – and lost it

Say No to Online Age Verification

Why implementing ActivityPub is hard, and why it doesn't have to be

The SRE Guide to Kubernetes Observability: Red vs. Use Methods

AI saves about 3% of your hours, and almost none of it reaches the money

I Wasn't Allowed Prompting ChatGPT During My Chalk Talk: This Is Discrimination

Musk's X poses "serious risk to Americans' privacy," advocates warn FTC

Trump has made more than $1B from crypto in a year

Instagram running ads promoting child sexual abuse material in India, BBC finds

BitTorrent Explained: Deep Dive into Peer-to-Peer Sharing, Security, and Usage

Cache Merging as a Convergent Replicated State for Multi-Agent Latent Reasoning

AI's $2.2T deficit fix is already half fake, economists say

Heuristics for lab robotics, and where its future may go

Costco Is the Anti-Amazon

Markets are competitive if and only if P != NP

International chess federation sanctions Kramnik

Evaluation order and nontermination in query languages

Minimum Connection Times at Airports Worldwide

3D Printing a Brain (2021)

What we learned building a multi-agent PDF table extractor