Hasty Briefsbeta

Bilingual

Dithering with CSS
4 days ago
- Using CSS for dithering allows websites to maintain a consistent aesthetic and color scheme across all pictures.
- While pre-processing and saving images already dithered might be more efficient, CSS dithering offers greater customization options.
ClojureScript Gets Async/Await
3 hours ago
- ClojureScript now supports JavaScript async functions through the ^:async hint.
- Async functions can use the await keyword to handle promises directly.
- The enhancement aims to improve interoperability with modern Browser APIs and libraries.
- This release includes updates targeting ECMAScript 2016.
HantaWatch Real time hantavirus outbreak tracker
5 hours ago
- A stewardess briefly on a Johannesburg-Amsterdam flight was admitted to Amsterdam UMC on May 7; tests are expected soon, marking a potential secondary case outside the MV Hondius ship.
- WHO confirms 8 total cases on the ship, with no new symptomatic passengers; 3 are lab-confirmed. The ship is en route to Spain (3-4 days), and global risk is assessed as LOW.
- Around 40 passengers have disembarked and are being tracked across the U.S. and 12+ international locations; contact tracing is ongoing.
- The Andes hantavirus strain is the only one capable of human-to-human transmission via respiratory droplets during close, prolonged contact, primarily in South America.
- Transmission window is short (~24 hours), with peak infectiousness on fever onset; even brief contact can transmit. Person-to-person spread is rare (2-5% historically).
- Case fatality rate is 30-50%; in the MV Hondius outbreak, 3 out of 8 cases died. Early hospital treatment with supportive care significantly improves survival.
- No vaccine exists; treatment is supportive (oxygen therapy, fluids). Antiviral ribavirin may help if given early. Early ICU hospitalization dramatically boosts survival rates.
- WHO emphasizes this is not COVID-19 or influenza; Andes spreads differently, requiring close contact, and there's no evidence of increased transmissibility from mutation.
- Recommendations include monitoring for fever/symptoms (incubation: 1-6 weeks), seeking immediate medical care if fever develops, and avoiding close contact while monitoring.
Why I'm giving up on Facebook (2017)
3 hours ago
- The author's New Year's resolution is to be more intentional about life, leading to the decision to quit Facebook after two months of logging out, similar to giving up sugar.
- Reasons for leaving Facebook include manipulative user interface designed to waste time, extensive data collection and sharing, and the creation of social media bubbles and echo chambers.
- Facebook enables data manipulation by third parties, isolates groups, and there are better ways to stay in touch, such as phone calls or in-person chats.
- The author compares Facebook to refined sugar—ingrained and hard to give up but ultimately negative—and is not deleting the account due to existing data retention and potential future needs.
- The author encourages real friendships and provides a Twitter handle for alternative contact, wishing everyone a happy and intentional 2017.
Silver, Silk, and States: The Spanish Empire and Ming China
8 hours ago
- The global silver trade in the 16th-17th centuries linked Habsburg Spain and Ming China into a single interdependent 'war-finance machine'.
- Ming China's Single-Whip Tax Reform created massive demand for silver, monetizing taxes and funding military professionalization.
- Spanish Empire's coerced silver production in Potosí and Zacatecas, via mita labor and mercury amalgamation, supplied the silver.
- The Manila Galleon trade route facilitated silver flow to China, driven by bimetallic arbitrage profits.
- Spanish silver funded Genoese credit for European wars, while Ming silver paid for armies defending frontiers.
- Systemic interdependence created shared vulnerability; silver supply disruptions led to crises in both empires.
- 17th-century silver supply shocks from Potosí decline, Spanish restrictions, and Japanese sakoku caused deflation and fiscal collapse in Ming China.
- The silver crisis accelerated Ming collapse (1644) and ended Spain's Golden Age, highlighting global integration's fragility.
Google's next-gen reCAPTCHA system could spell trouble for de-Googled phones
13 hours ago
- Google's next-gen reCAPTCHA system mandates Google Play Services version 25.41.30 or higher for Android devices.
- De-Googled phones (e.g., those with GrapheneOS) will fail verification if reCAPTCHA escalates to QR code scanning.
- The change, in development since October last year, could hinder web browsing for users without Google Play Services.
DBase: 1979-2026
5 hours ago
- dBase, once the dominant database for IBM-compatible PCs from 1979, faced decline due to litigation and neglect, with its newsgroups going offline in November 2025.
- Ashton-Tate's lawsuits and piracy audits in the legacy era (1979-1995) chilled the developer community, while competitors like FoxPro and Btrieve addressed customer demands for network-aware servers.
- During the dBase for Windows era (1995-2012), compatibility issues persisted, and the source code was lost after the Borland-Corel merger, though the file format remained in use for GIS and mainframes.
- From 2012-2019, dBase LLC released updates but retained the outdated BDE runtime; development stalled, and third-party vendors trapped customers with binary add-ons and non-existent upgrades.
- Recent AI advancements enable migration of legacy dBase code (e.g., .PRG files) to modern languages like Rust, Go, or Dart/Flutter, offering an escape route for those still in the ecosystem.
KDE Plasma 6.7 to Provide a Better Experience for CPU-Based Rendering
4 hours ago
- KDE Plasma 6.7 introduces UDMABUF for CPU-based rendering improvements.
- UDMABUF reduces excess buffer copies in Wayland shared memory usage.
- KWin will create UDMABUF for each wl_shm buffer, optimizing GPU import.
- Performance gains include smoother cursor movement and lower CPU usage in applications like KDevelop.
Digging into Drama at the Document Foundation
6 hours ago
- The Document Foundation (TDF) revoked membership of about 30 people affiliated with Collabora, citing concerns over maintaining its nonprofit status.
- Conflicts of interest were identified, including past board decisions benefiting companies like Collabora and CIB through contracts and trademark usage.
- Audits revealed potential misuse of funds and governance issues, leading TDF to adopt new bylaws allowing membership revocation for legal disputes.
- Collabora announced a shift away from LibreOffice toward its own 'Collabora Office' project, reducing involvement with TDF.
- The membership committee delayed board elections and faced internal disputes over authority and compliance with statutes.
- TDF aims to hire more developers and seek new partnerships to offset potential reduced contributions from Collabora.
- Ongoing mediation and governance tensions persist, with elections and future strategic planning pending.
He Couldn't Land a Job Interview. Was AI to Blame?
16 hours ago
- Chad Markey, a medical student with strong credentials, faced unexpected rejections and no interview offers for residencies, suspecting an AI screening tool called Cortex might be to blame.
- Markey's Medical Student Performance Evaluation (MSPE) mentioned 'voluntary' leaves of absence for personal reasons, which he felt misrepresented his medically necessary absences due to ankylosing spondylitis, potentially triggering AI bias.
- He embarked on a six-month investigation, coding and analyzing how AI might interpret his application, including creating synthetic datasets and reverse-engineering a patent for a similar AI tool.
- After manually emailing programs about a new research publication, Markey received immediate interview offers, leading him to believe his application was initially overlooked by AI-driven screening.
- Issues with Cortex were reported, including inaccurate grade displays, though Thalamus, the maker, stated these were minimal and corrected, and clarified that Cortex does not use AI to score or rank applicants.
- Markey successfully matched with Columbia University's psychiatry program but continued his research into AI bias, highlighting the lack of transparency and regulation in AI hiring tools.
- Background-check AI tools are regulated under the Fair Credit Reporting Act, offering some recourse, unlike many AI screening systems in hiring, which lack similar oversight for individual complaints.
New open source city-state, with new constitution functional on one site
9 hours ago
- Warning: Lying about UBI, benefits, or taxes can result in fines up to 10 times the amount and up to 1 year imprisonment, with a jury trial.
- Reference to the Declaration of Independence emphasizing equality, unalienable rights, and the right to alter government.
- Citizen Portal access details: Women receive maternity pay per child under 3, starting at $2,400/month at birth and tapering over 3 years, totaling $36,000 per child.
- UBI and benefits are paid monthly to a Bitcoin Cash address, adjusted by a percentage to balance Arkology's budget; displayed figures are maximums.
- Taxes must be paid within 90 days or incur a 30% fee; failure after 180 days incurs a 90% fee, possible court action, and expulsion from Arkology.
The IDE Should Become an Operating System for AI
9 hours ago
- Modern IDEs have limitations for AI work due to their file and panel-based organization, lacking flexibility for integrating diverse components like terminals, logs, and agents.
- Historical systems like Smalltalk, Lisp machines, and Emacs provided more flexible, inspectable environments by treating the running system as an addressable object space or using buffers as a universal substrate.
- For AI to thrive, every element—such as file lines, terminal byte ranges, diffs, or network requests—should have an identity, making them addressable, inspectable, searchable, and replayable.
- Buffers should be redefined as durable work items with types, URIs, and capabilities, applicable not just to text but also to terminals, tests, browser replays, and agent runs, with consistent commands across them.
- The future IDE needs to evolve into an AI operating system, featuring a scheduler, memory, object model, permissions, event logs, and a command language that serves both humans and agents, ensuring safety and coherence without hidden UI powers.
- This shift transforms the IDE from a code-typing tool into a runtime for software work, making Git, browsers, terminals, and deployments coherent objects in a unified space navigable by users and AI.
GPT-5.5 Price Increase: What It Costs
9 hours ago
- GPT-5.5 launched with a 2x price increase over GPT-5.4, with input tokens rising from $2.50/M to $5.00/M and output tokens from $15/M to $30/M.
- Analysis using a 'switcher cohort' of users who shifted from GPT-5.4 to GPT-5.5 showed cost increases between 49% and 92%, depending on prompt size.
- GPT-5.5 is less verbose for longer prompts (over 10K tokens), producing 19-34% fewer completion tokens, which partially offsets the price hike.
- For shorter prompts (under 10K tokens), completion lengths increased or stayed similar, leading to higher cost increases, up to 92% for the smallest prompts.
- Methodology involved OpenRouter's request logs, focusing on text-only, non-cancelled requests, with consistent token counts and normalization to cost per million tokens.
Evaluating Geekbench 6
7 hours ago
- Geekbench 6 is a consumer-focused benchmark suite distributed in binary form, unlike the source-code-based SPEC CPU2017, and is evaluated alongside SPEC to understand workload challenges for modern CPUs.
- Geekbench 6 heavily utilizes vector instructions, with AVX(2) prevalent in most workloads and AVX-512 prominent in specific tests like Background Blur; AMX is also used minimally but effectively in Object Detection and Photo Library.
- Instruction per cycle (IPC) analysis shows Geekbench 6 workloads skew towards medium to high IPC, indicating good core throughput, while SPEC CPU2017 has a wider IPC distribution with more low-IPC challenges, especially in integer tests.
- Branch prediction is critical, with Navigation being a particularly challenging workload due to high mispredicts, limiting IPC gains even on modern cores, while other tests like File Compression and Clang present moderate challenges.
- Cache and memory access patterns show Geekbench 6 workloads often have good data locality, with low L3 misses and effective prefetching, though tests like Photo Filter generate high DRAM bandwidth despite low miss rates.
- Score calculation differs: Geekbench 6 uses a modern reference system (Core i7-12700) with a baseline of 2500, while SPEC CPU2017 uses an older reference (Sun Fire V490) with a score of 1 as a speedup ratio.
- Overall, Geekbench 6 emphasizes vector-heavy, throughput-focused workloads with small instruction footprints, whereas SPEC CPU2017 offers broader IPC challenges but less vector stress, serving complementary purposes in benchmarking.
US will start revoking passports for parents who owe child support
6 hours ago
- The U.S. State Department will start revoking passports for parents owing significant unpaid child support.
- Initial focus is on those owing $100,000 or more (about 2,700 passport holders), starting Friday.
- The program will soon expand to cover parents owing over $2,500, as per a 1996 law.
- Previously, only passport renewals triggered revocation; now, all eligible passport holders will be targeted proactively.
- HHS provides data to the State Department, and revocations aim to encourage debt resolution.
- Since February news of the expansion, hundreds of parents resolved arrears, showing the policy's effectiveness.
- The program, active since 1998, has helped states collect $657 million, including $156 million in lump sums over five years.
- Revoked passport holders are notified and must apply for new passports after paying arrears.
- Those abroad during revocation must visit U.S. embassies/consulates for emergency travel documents to return.
A polynomial autoencoder beats PCA on transformer embeddings
3 days ago
- Polynomial autoencoder improves compression of embeddings by adding a quadratic decoder to PCA, capturing nonlinear structure missed by linear methods.
- The method uses a closed-form quadratic lift with Ridge OLS regression, eliminating the need for SGD or hyperparameter tuning.
- Experiments on BEIR/FiQA show poly-AE reduces the performance gap to raw embeddings significantly, achieving 4× compression with minimal NDCG loss.
- Poly-AE consistently outperforms PCA, especially at higher compression levels (e.g., d=128) and on non-matryoshka-trained models.
- Limitations include cubic computational cost for large d, transductive fitting requiring corpus statistics, and overfitting on small corpora.
- The technique originates from quadratic manifold methods in dynamical systems and is adapted here for neural embeddings.
- Future directions include testing on larger datasets, exploring higher-degree polynomials, and hybrid approaches with matryoshka embeddings.
CERN's KiCad component library now open source
16 hours ago
- CERN has released its complete KiCad component library as open source, containing over 17,000 electronic components.
- KiCad is a free, open-source tool for PCB design, first released in 1992, enabling unrestricted sharing and collaboration.
- CERN's release supports the open-source philosophy, allowing study, modification, redistribution, and commercialization of hardware designs.
- This move complements CERN's broader open science initiatives, including open hardware licensing, software, access publishing, and data.
How to make SSE token streams resumable, cancellable, and multi-device
2 days ago
- Agents have evolved from synchronous interactions to background operations, breaking traditional transport methods.
- Advanced chatbot features include resumable streams, cancellations, and multi-device support, achievable but not necessarily easy with Server-Sent Events (SSE).
- LLM responses consist of tokens with metadata; storing each token for resumability leads to inefficient database writes and cleanup.
- Resumable streams require storing tokens in a shared database due to stateless server replicas, increasing write amplification.
- Cancellations require a separate endpoint and shared store to signal abort, complicating dropped connection handling.
- Multi-device support involves sharing token streams and real-time updates, often necessitating polling or long-polling solutions.
- SSE over HTTP is criticized as inefficient for streaming LLM tokens; pub/sub patterns offer better decoupling and automation for AI applications.
Why Don't Lowercase Letters Come Right After Uppercase Letters in ASCII?
12 hours ago
- ASCII uses 7 bits, allowing 128 code points, and its first 128 code points are the same as Unicode.
- There are six characters between uppercase Z and lowercase a, making 26 + 6 = 32, a power of two.
- Uppercase and lowercase letters differ by the 5th bit, corresponding to decimal 32.
- Bitwise operations like AND, OR, and XOR with 32 allow easy case conversion and flipping.
- Masking with 31 (binary 00011111) extracts alphabet position, as c & 31 equals c % 32.
Anthropic response to 1-click pwn: Shouldn't have clicked 'ok'
7 hours ago
- MOST POPULAR section highlights key events and articles, including topics on AI security, supply chain issues, and identity resilience.
- EVENTS include workshops on securing AI development, toxic AI agent skills, hardware supply chain turbulence, and identity-first recovery strategies.
- AI section covers self-driving Wi-Fi collaboration, AI-assisted bug detection in Firefox, and AI tools for ServiceNow configurations.
- Cyber-crime reports detail prison sentences for IT fraud involving North Korean scammers and warnings about AI tool security.
- FOSS features emphasize open-source security issues, AI tool effectiveness, and the persistence of vulnerable password hashes like MD5.
- FEATURES explore topics such as GNOME vs X.org, containerized applications, open-source security funding, and the impact of AI on memory systems.

first prev0next

About|Login

Bilingual

Dithering with CSS
4 days ago
- Using CSS for dithering allows websites to maintain a consistent aesthetic and color scheme across all pictures.
- While pre-processing and saving images already dithered might be more efficient, CSS dithering offers greater customization options.
ClojureScript Gets Async/Await
3 hours ago
- ClojureScript now supports JavaScript async functions through the ^:async hint.
- Async functions can use the await keyword to handle promises directly.
- The enhancement aims to improve interoperability with modern Browser APIs and libraries.
- This release includes updates targeting ECMAScript 2016.
HantaWatch Real time hantavirus outbreak tracker
5 hours ago
- A stewardess briefly on a Johannesburg-Amsterdam flight was admitted to Amsterdam UMC on May 7; tests are expected soon, marking a potential secondary case outside the MV Hondius ship.
- WHO confirms 8 total cases on the ship, with no new symptomatic passengers; 3 are lab-confirmed. The ship is en route to Spain (3-4 days), and global risk is assessed as LOW.
- Around 40 passengers have disembarked and are being tracked across the U.S. and 12+ international locations; contact tracing is ongoing.
- The Andes hantavirus strain is the only one capable of human-to-human transmission via respiratory droplets during close, prolonged contact, primarily in South America.
- Transmission window is short (~24 hours), with peak infectiousness on fever onset; even brief contact can transmit. Person-to-person spread is rare (2-5% historically).
- Case fatality rate is 30-50%; in the MV Hondius outbreak, 3 out of 8 cases died. Early hospital treatment with supportive care significantly improves survival.
- No vaccine exists; treatment is supportive (oxygen therapy, fluids). Antiviral ribavirin may help if given early. Early ICU hospitalization dramatically boosts survival rates.
- WHO emphasizes this is not COVID-19 or influenza; Andes spreads differently, requiring close contact, and there's no evidence of increased transmissibility from mutation.
- Recommendations include monitoring for fever/symptoms (incubation: 1-6 weeks), seeking immediate medical care if fever develops, and avoiding close contact while monitoring.
Why I'm giving up on Facebook (2017)
3 hours ago
- The author's New Year's resolution is to be more intentional about life, leading to the decision to quit Facebook after two months of logging out, similar to giving up sugar.
- Reasons for leaving Facebook include manipulative user interface designed to waste time, extensive data collection and sharing, and the creation of social media bubbles and echo chambers.
- Facebook enables data manipulation by third parties, isolates groups, and there are better ways to stay in touch, such as phone calls or in-person chats.
- The author compares Facebook to refined sugar—ingrained and hard to give up but ultimately negative—and is not deleting the account due to existing data retention and potential future needs.
- The author encourages real friendships and provides a Twitter handle for alternative contact, wishing everyone a happy and intentional 2017.
Silver, Silk, and States: The Spanish Empire and Ming China
8 hours ago
- The global silver trade in the 16th-17th centuries linked Habsburg Spain and Ming China into a single interdependent 'war-finance machine'.
- Ming China's Single-Whip Tax Reform created massive demand for silver, monetizing taxes and funding military professionalization.
- Spanish Empire's coerced silver production in Potosí and Zacatecas, via mita labor and mercury amalgamation, supplied the silver.
- The Manila Galleon trade route facilitated silver flow to China, driven by bimetallic arbitrage profits.
- Spanish silver funded Genoese credit for European wars, while Ming silver paid for armies defending frontiers.
- Systemic interdependence created shared vulnerability; silver supply disruptions led to crises in both empires.
- 17th-century silver supply shocks from Potosí decline, Spanish restrictions, and Japanese sakoku caused deflation and fiscal collapse in Ming China.
- The silver crisis accelerated Ming collapse (1644) and ended Spain's Golden Age, highlighting global integration's fragility.
Google's next-gen reCAPTCHA system could spell trouble for de-Googled phones
13 hours ago
- Google's next-gen reCAPTCHA system mandates Google Play Services version 25.41.30 or higher for Android devices.
- De-Googled phones (e.g., those with GrapheneOS) will fail verification if reCAPTCHA escalates to QR code scanning.
- The change, in development since October last year, could hinder web browsing for users without Google Play Services.
DBase: 1979-2026
5 hours ago
- dBase, once the dominant database for IBM-compatible PCs from 1979, faced decline due to litigation and neglect, with its newsgroups going offline in November 2025.
- Ashton-Tate's lawsuits and piracy audits in the legacy era (1979-1995) chilled the developer community, while competitors like FoxPro and Btrieve addressed customer demands for network-aware servers.
- During the dBase for Windows era (1995-2012), compatibility issues persisted, and the source code was lost after the Borland-Corel merger, though the file format remained in use for GIS and mainframes.
- From 2012-2019, dBase LLC released updates but retained the outdated BDE runtime; development stalled, and third-party vendors trapped customers with binary add-ons and non-existent upgrades.
- Recent AI advancements enable migration of legacy dBase code (e.g., .PRG files) to modern languages like Rust, Go, or Dart/Flutter, offering an escape route for those still in the ecosystem.
KDE Plasma 6.7 to Provide a Better Experience for CPU-Based Rendering
4 hours ago
- KDE Plasma 6.7 introduces UDMABUF for CPU-based rendering improvements.
- UDMABUF reduces excess buffer copies in Wayland shared memory usage.
- KWin will create UDMABUF for each wl_shm buffer, optimizing GPU import.
- Performance gains include smoother cursor movement and lower CPU usage in applications like KDevelop.
Digging into Drama at the Document Foundation
6 hours ago
- The Document Foundation (TDF) revoked membership of about 30 people affiliated with Collabora, citing concerns over maintaining its nonprofit status.
- Conflicts of interest were identified, including past board decisions benefiting companies like Collabora and CIB through contracts and trademark usage.
- Audits revealed potential misuse of funds and governance issues, leading TDF to adopt new bylaws allowing membership revocation for legal disputes.
- Collabora announced a shift away from LibreOffice toward its own 'Collabora Office' project, reducing involvement with TDF.
- The membership committee delayed board elections and faced internal disputes over authority and compliance with statutes.
- TDF aims to hire more developers and seek new partnerships to offset potential reduced contributions from Collabora.
- Ongoing mediation and governance tensions persist, with elections and future strategic planning pending.
He Couldn't Land a Job Interview. Was AI to Blame?
16 hours ago
- Chad Markey, a medical student with strong credentials, faced unexpected rejections and no interview offers for residencies, suspecting an AI screening tool called Cortex might be to blame.
- Markey's Medical Student Performance Evaluation (MSPE) mentioned 'voluntary' leaves of absence for personal reasons, which he felt misrepresented his medically necessary absences due to ankylosing spondylitis, potentially triggering AI bias.
- He embarked on a six-month investigation, coding and analyzing how AI might interpret his application, including creating synthetic datasets and reverse-engineering a patent for a similar AI tool.
- After manually emailing programs about a new research publication, Markey received immediate interview offers, leading him to believe his application was initially overlooked by AI-driven screening.
- Issues with Cortex were reported, including inaccurate grade displays, though Thalamus, the maker, stated these were minimal and corrected, and clarified that Cortex does not use AI to score or rank applicants.
- Markey successfully matched with Columbia University's psychiatry program but continued his research into AI bias, highlighting the lack of transparency and regulation in AI hiring tools.
- Background-check AI tools are regulated under the Fair Credit Reporting Act, offering some recourse, unlike many AI screening systems in hiring, which lack similar oversight for individual complaints.
New open source city-state, with new constitution functional on one site
9 hours ago
- Warning: Lying about UBI, benefits, or taxes can result in fines up to 10 times the amount and up to 1 year imprisonment, with a jury trial.
- Reference to the Declaration of Independence emphasizing equality, unalienable rights, and the right to alter government.
- Citizen Portal access details: Women receive maternity pay per child under 3, starting at $2,400/month at birth and tapering over 3 years, totaling $36,000 per child.
- UBI and benefits are paid monthly to a Bitcoin Cash address, adjusted by a percentage to balance Arkology's budget; displayed figures are maximums.
- Taxes must be paid within 90 days or incur a 30% fee; failure after 180 days incurs a 90% fee, possible court action, and expulsion from Arkology.
The IDE Should Become an Operating System for AI
9 hours ago
- Modern IDEs have limitations for AI work due to their file and panel-based organization, lacking flexibility for integrating diverse components like terminals, logs, and agents.
- Historical systems like Smalltalk, Lisp machines, and Emacs provided more flexible, inspectable environments by treating the running system as an addressable object space or using buffers as a universal substrate.
- For AI to thrive, every element—such as file lines, terminal byte ranges, diffs, or network requests—should have an identity, making them addressable, inspectable, searchable, and replayable.
- Buffers should be redefined as durable work items with types, URIs, and capabilities, applicable not just to text but also to terminals, tests, browser replays, and agent runs, with consistent commands across them.
- The future IDE needs to evolve into an AI operating system, featuring a scheduler, memory, object model, permissions, event logs, and a command language that serves both humans and agents, ensuring safety and coherence without hidden UI powers.
- This shift transforms the IDE from a code-typing tool into a runtime for software work, making Git, browsers, terminals, and deployments coherent objects in a unified space navigable by users and AI.
GPT-5.5 Price Increase: What It Costs
9 hours ago
- GPT-5.5 launched with a 2x price increase over GPT-5.4, with input tokens rising from $2.50/M to $5.00/M and output tokens from $15/M to $30/M.
- Analysis using a 'switcher cohort' of users who shifted from GPT-5.4 to GPT-5.5 showed cost increases between 49% and 92%, depending on prompt size.
- GPT-5.5 is less verbose for longer prompts (over 10K tokens), producing 19-34% fewer completion tokens, which partially offsets the price hike.
- For shorter prompts (under 10K tokens), completion lengths increased or stayed similar, leading to higher cost increases, up to 92% for the smallest prompts.
- Methodology involved OpenRouter's request logs, focusing on text-only, non-cancelled requests, with consistent token counts and normalization to cost per million tokens.
Evaluating Geekbench 6
7 hours ago
- Geekbench 6 is a consumer-focused benchmark suite distributed in binary form, unlike the source-code-based SPEC CPU2017, and is evaluated alongside SPEC to understand workload challenges for modern CPUs.
- Geekbench 6 heavily utilizes vector instructions, with AVX(2) prevalent in most workloads and AVX-512 prominent in specific tests like Background Blur; AMX is also used minimally but effectively in Object Detection and Photo Library.
- Instruction per cycle (IPC) analysis shows Geekbench 6 workloads skew towards medium to high IPC, indicating good core throughput, while SPEC CPU2017 has a wider IPC distribution with more low-IPC challenges, especially in integer tests.
- Branch prediction is critical, with Navigation being a particularly challenging workload due to high mispredicts, limiting IPC gains even on modern cores, while other tests like File Compression and Clang present moderate challenges.
- Cache and memory access patterns show Geekbench 6 workloads often have good data locality, with low L3 misses and effective prefetching, though tests like Photo Filter generate high DRAM bandwidth despite low miss rates.
- Score calculation differs: Geekbench 6 uses a modern reference system (Core i7-12700) with a baseline of 2500, while SPEC CPU2017 uses an older reference (Sun Fire V490) with a score of 1 as a speedup ratio.
- Overall, Geekbench 6 emphasizes vector-heavy, throughput-focused workloads with small instruction footprints, whereas SPEC CPU2017 offers broader IPC challenges but less vector stress, serving complementary purposes in benchmarking.
US will start revoking passports for parents who owe child support
6 hours ago
- The U.S. State Department will start revoking passports for parents owing significant unpaid child support.
- Initial focus is on those owing $100,000 or more (about 2,700 passport holders), starting Friday.
- The program will soon expand to cover parents owing over $2,500, as per a 1996 law.
- Previously, only passport renewals triggered revocation; now, all eligible passport holders will be targeted proactively.
- HHS provides data to the State Department, and revocations aim to encourage debt resolution.
- Since February news of the expansion, hundreds of parents resolved arrears, showing the policy's effectiveness.
- The program, active since 1998, has helped states collect $657 million, including $156 million in lump sums over five years.
- Revoked passport holders are notified and must apply for new passports after paying arrears.
- Those abroad during revocation must visit U.S. embassies/consulates for emergency travel documents to return.
A polynomial autoencoder beats PCA on transformer embeddings
3 days ago
- Polynomial autoencoder improves compression of embeddings by adding a quadratic decoder to PCA, capturing nonlinear structure missed by linear methods.
- The method uses a closed-form quadratic lift with Ridge OLS regression, eliminating the need for SGD or hyperparameter tuning.
- Experiments on BEIR/FiQA show poly-AE reduces the performance gap to raw embeddings significantly, achieving 4× compression with minimal NDCG loss.
- Poly-AE consistently outperforms PCA, especially at higher compression levels (e.g., d=128) and on non-matryoshka-trained models.
- Limitations include cubic computational cost for large d, transductive fitting requiring corpus statistics, and overfitting on small corpora.
- The technique originates from quadratic manifold methods in dynamical systems and is adapted here for neural embeddings.
- Future directions include testing on larger datasets, exploring higher-degree polynomials, and hybrid approaches with matryoshka embeddings.
CERN's KiCad component library now open source
16 hours ago
- CERN has released its complete KiCad component library as open source, containing over 17,000 electronic components.
- KiCad is a free, open-source tool for PCB design, first released in 1992, enabling unrestricted sharing and collaboration.
- CERN's release supports the open-source philosophy, allowing study, modification, redistribution, and commercialization of hardware designs.
- This move complements CERN's broader open science initiatives, including open hardware licensing, software, access publishing, and data.
How to make SSE token streams resumable, cancellable, and multi-device
2 days ago
- Agents have evolved from synchronous interactions to background operations, breaking traditional transport methods.
- Advanced chatbot features include resumable streams, cancellations, and multi-device support, achievable but not necessarily easy with Server-Sent Events (SSE).
- LLM responses consist of tokens with metadata; storing each token for resumability leads to inefficient database writes and cleanup.
- Resumable streams require storing tokens in a shared database due to stateless server replicas, increasing write amplification.
- Cancellations require a separate endpoint and shared store to signal abort, complicating dropped connection handling.
- Multi-device support involves sharing token streams and real-time updates, often necessitating polling or long-polling solutions.
- SSE over HTTP is criticized as inefficient for streaming LLM tokens; pub/sub patterns offer better decoupling and automation for AI applications.
Why Don't Lowercase Letters Come Right After Uppercase Letters in ASCII?
12 hours ago
- ASCII uses 7 bits, allowing 128 code points, and its first 128 code points are the same as Unicode.
- There are six characters between uppercase Z and lowercase a, making 26 + 6 = 32, a power of two.
- Uppercase and lowercase letters differ by the 5th bit, corresponding to decimal 32.
- Bitwise operations like AND, OR, and XOR with 32 allow easy case conversion and flipping.
- Masking with 31 (binary 00011111) extracts alphabet position, as c & 31 equals c % 32.
Anthropic response to 1-click pwn: Shouldn't have clicked 'ok'
7 hours ago
- MOST POPULAR section highlights key events and articles, including topics on AI security, supply chain issues, and identity resilience.
- EVENTS include workshops on securing AI development, toxic AI agent skills, hardware supply chain turbulence, and identity-first recovery strategies.
- AI section covers self-driving Wi-Fi collaboration, AI-assisted bug detection in Firefox, and AI tools for ServiceNow configurations.
- Cyber-crime reports detail prison sentences for IT fraud involving North Korean scammers and warnings about AI tool security.
- FOSS features emphasize open-source security issues, AI tool effectiveness, and the persistence of vulnerable password hashes like MD5.
- FEATURES explore topics such as GNOME vs X.org, containerized applications, open-source security funding, and the impact of AI on memory systems.

Hasty Briefsbeta

Dithering with CSS

ClojureScript Gets Async/Await

HantaWatch Real time hantavirus outbreak tracker

Why I'm giving up on Facebook (2017)

Silver, Silk, and States: The Spanish Empire and Ming China

Google's next-gen reCAPTCHA system could spell trouble for de-Googled phones

DBase: 1979-2026

KDE Plasma 6.7 to Provide a Better Experience for CPU-Based Rendering

Digging into Drama at the Document Foundation

He Couldn't Land a Job Interview. Was AI to Blame?

New open source city-state, with new constitution functional on one site

The IDE Should Become an Operating System for AI

GPT-5.5 Price Increase: What It Costs

Evaluating Geekbench 6

US will start revoking passports for parents who owe child support

A polynomial autoencoder beats PCA on transformer embeddings

CERN's KiCad component library now open source

How to make SSE token streams resumable, cancellable, and multi-device

Why Don't Lowercase Letters Come Right After Uppercase Letters in ASCII?

Anthropic response to 1-click pwn: Shouldn't have clicked 'ok'

Dithering with CSS

ClojureScript Gets Async/Await

HantaWatch Real time hantavirus outbreak tracker

Why I'm giving up on Facebook (2017)

Silver, Silk, and States: The Spanish Empire and Ming China

Google's next-gen reCAPTCHA system could spell trouble for de-Googled phones

DBase: 1979-2026

KDE Plasma 6.7 to Provide a Better Experience for CPU-Based Rendering

Digging into Drama at the Document Foundation

He Couldn't Land a Job Interview. Was AI to Blame?

New open source city-state, with new constitution functional on one site

The IDE Should Become an Operating System for AI

GPT-5.5 Price Increase: What It Costs

Evaluating Geekbench 6

US will start revoking passports for parents who owe child support

A polynomial autoencoder beats PCA on transformer embeddings

CERN's KiCad component library now open source

How to make SSE token streams resumable, cancellable, and multi-device

Why Don't Lowercase Letters Come Right After Uppercase Letters in ASCII?

Anthropic response to 1-click pwn: Shouldn't have clicked 'ok'