Hasty Briefsbeta

Bilingual

Adobe Has Run Out of Allies
4 hours ago
- Adobe shifted from Creative Suite to Creative Cloud in 2013, promising more agility, updates, and affordability.
- Initially, Adobe's developers and product managers actively engaged with media and users, showing care and pride.
- Adobe's focus shifted from individual creative users to corporate clients, altering its media strategy and communication.
- The company replaced customer-facing teams with business-oriented managers and legal messaging, alienating artists.
- Adobe now faces competition from Apple, Blackmagic, and Canva, with declining stock value and user loyalty.
- Individual users and shareholders have turned against Adobe, leaving the company isolated and widely detested.
Hardware Is Hard?
4 hours ago
- The author, feeling bored with AI taking over their work, decides to build a robot for fun, inspired by nostalgic college projects.
- Researching robot building reveals a lack of electronics knowledge, including basics like electricity, components, and tools such as multimeters and breadboards.
- To start simply, the author chooses to build a "Hello, World!" equivalent in robotics: a small remote-controlled robot car using a phone.
- The core principle is that voltage difference moves motors, explained through Ohm's law (V=IR) with an analogy to water flowing downhill.
- Required components include a 4WD chassis kit, 18650 Li-ion batteries, ESP32 as the brain, TB6612FNG motor controller, multimeter, and other tools like a soldering iron.
- Assembly involves testing motor directions, wiring components, and using the ESP32 to control motors via GPIO pins and pulse width modulation (PWM) for speed.
- The TB6612FNG module acts as an H-bridge to control motor direction and speed based on ESP32 signals, avoiding energy waste from resistance.
- Programming the ESP32 involves C code interacting with memory-mapped I/O to trigger pins, with help from online resources and AI like Claude.
- Wiring follows detailed instructions, and despite challenges like faulty components, the author succeeds in making the wheels move, achieving a eureka moment.
- The author concludes that electronics revolves around voltage, appreciates AI assistance, and encourages embracing life's richness through hands-on projects.
PostgreSQL production incident caused by transaction ID wraparound
3 hours ago
- PostgreSQL transaction ID wraparound is a silent but severe failure mode that can cause a complete write outage.
- It occurs when freezing of old row versions doesn't happen, allowing transaction IDs to approach a hard safety limit.
- The failure is not immediate; it can manifest after months or years of normal operation with no warning signs.
- In the described incident, disabling autovacuum on tables, including unused ones, led to transaction ID aging.
- Once the limit is reached, PostgreSQL intentionally blocks all writes to prevent data corruption, making the database read-only.
- Recovery requires manual VACUUM FREEZE operations to advance transaction IDs beyond the danger zone.
- Transaction ID consumption depends on write rate and time; even modest workloads can lead to risk after about 7.5 months.
- Unlike PostgreSQL, SQL Server avoids this issue by using Log Sequence Numbers (LSNs) instead of finite, reusable transaction IDs.
- Understanding transaction ID math and treating autovacuum as a safety mechanism is crucial for PostgreSQL production stability.
Optimizing Ruby Path Methods
3 hours ago
- Bootsnap optimizes Ruby's file loading by caching load paths to avoid repeated expensive searches.
- Load path caching converts O(N*M) boot time complexity into O(1) lookups, speeding up application startup.
- Cache invalidation relies on directory mtime, but git doesn't update mtimes, requiring frequent cache rebuilds on CI.
- Bootsnap's path scanner had an N+1 syscall issue due to stat calls for each directory entry.
- Dir.scan was introduced in Ruby 4.1 to expose file types without extra stat calls, improving scanning speed by 2x.
- File.join was optimized by implementing a fast path for ASCII-compatible encodings and fixing inefficient reverse search.
- Encoding handling, especially for Shift JIS, previously slowed File.join; a fast path now speeds up UTF-8 and ASCII cases.
- Optimizations to File.join made it over 7 times faster for common cases, outperforming simple string interpolation.
- Other path methods like File.basename, File.extname, and File.dirname received similar optimizations for consistency.
I learned Unity the wrong way
3 days ago
- The author shares their experience of learning Unity in an inefficient or incorrect manner
- They reflect on common pitfalls and mistakes made during the learning process
- The article likely provides advice on better approaches to learning Unity effectively
- It emphasizes human-written content focused on personal development insights
Shader Lab, like Photoshop but for shaders
2 days ago
- The text appears to be a UI settings panel for a CRT (Cathode Ray Tube) effect, likely from a video editing or graphics software.
- It lists various properties and adjustment sliders under categories such as General, Settings, Distortion, Phosphor, Persistence, Noise, Glitch, and Bloom.
- Key adjustable parameters include Opacity, Blend Mode, Hue, Saturation, Mask Scale, Scanline Intensity, Mask Intensity, Barrel Distortion, Convergence, Vignette, Beam Focus, Brightness, Highlight Drive, Shoulder, Chroma Retention, Shadow Lift, Flicker, Glitch Speed, and Bloom settings (Intensity, Threshold, Radius, Softness).
IPC medley: message-queue peeking, io_uring, and bus1
2 days ago
- Proposal for mq_timedreceive2() adds flags and index parameters for peeking at message queues, useful for monitoring and checkpoint/restore.
- io_uring IPC patch series aims to create a high-bandwidth communication mechanism using shared ring buffers, similar to D-Bus, but needs more development.
- Bus1 returns rewritten in Rust, offering kernel-mediated IPC with message and capability passing, seeking integration help from Rust for Linux community.
Year of the IPv6 Overlay Network
a day ago
- Nebula 1.10 introduces IPv6 support in overlay networks, enabling future-proofing and seamless connectivity.
- A new v2 certificate format using ASN.1 replaces Protobufs, supporting IPv6 and multiple IPs per host.
- Multiple Nebula IPs per host can now be assigned, including both IPv4 and IPv6 addresses.
- IPv6 adoption eliminates addressing conflicts and prepares networks for IPv6 dominance.
- Upgrade process involves upgrading to v1.10, creating a v2 CA, and minting v2 certs with IPv4/IPv6 addresses.
- Firewall behavior changes: default_local_cidr_any now defaults to false for improved security.
- PKCS#11 support for P256 keys enables enterprise HSM and smart card integration.
- ECMP support for unsafe_routes provides load balancing and redundancy with multiple gateways.
- SO_MARK socket option support on Linux allows advanced routing via firewall marks.
- Enhanced library integration offers configurable logging and versioning for embedded use.
Comparing GPT-5.4, Opus 4.6, GLM-5.1, Kimi K2.5, MiMo V2 Pro and MiniMax M2.7
6 hours ago
- A benchmarking experiment compared six large language models (GPT-5.4, Opus 4.6, GLM-5.1, Kimi K2.5, MiMo V2 Pro, MiniMax M2.7) for building a native macOS app to manage drive ejection issues.
- All models performed similarly in planning and compilation phases, with MiMo V2 Pro notably using kernel API instead of shell commands, though this was considered irrelevant.
- Post-compilation, some apps crashed initially, but all could be fixed with comparable effort, highlighting a lack of end-to-end testing tools for native apps as a limitation.
- In code quality, GPT-5.4 and GLM-5.1 were preferred for leaner, more maintainable output, though prompting could guide any model toward cleaner code.
- A peer review ranking placed GPT-5.4 and Opus 4.6 consistently in the top three, with GPT-5.4 leading; MiMo V2 Pro also ranked high but scored lower on cleanliness, while GLM-5.1 excelled in cleanliness but placed fourth overall.
- Kimi K2.5 and MiniMax M2.7 consistently ranked at the bottom, correlating with their lower cost, and models showed little self-bias in scoring, except GPT-5.4, which ranked itself first, aligning with consensus.
- The conclusion recommends any top model for practical use, noting GLM and MiMo offer competitive performance at lower costs, and emphasizes picking one model to avoid procrastination.
The Mystery of Rennes-Le-Château, Part 4: Non-Fiction Meets Fiction
10 hours ago
- The Holy Blood and the Holy Grail was published in 1982, gaining strong sales and cultural traction on both sides of the Atlantic.
- Its success was influenced by timing, with Raiders of the Lost Ark's popularity and Britain's puzzle-book craze post-Masquerade.
- The U.S. paperback edition had an initial print run of 500,000 copies; translations and millions sold made the authors wealthy.
- Critics dismissed the book as shoddy history, but literary figures like Anthony Burgess praised its novelistic approach.
- Pierre Plantard distanced himself from the book's claim that Merovingian kings were descendants of Jesus Christ.
- Subsequent works like The Messianic Legacy (1987) saw declining impact, with Henry Lincoln obsessively revisiting theories.
- Graham Hancock emerged as a prominent figure in pseudo-history with books like Fingerprints of the Gods.
- Rennes-le-Château lore influenced media, including adventure games like Broken Sword and Gabriel Knight 3.
- The Tomb of God (1996) by Schellenberger and Andrews revived interest, proposing Jesus's body was hidden near Rennes-le-Château.
- Skeptics like Umberto Eco (Foucault's Pendulum) and BBC's Timewatch debunked many conspiracy claims.
- Dan Brown's The Da Vinci Code (2003) became a global phenomenon, borrowing heavily from the conspiracy theories.
- Leigh and Baigent sued Brown for plagiarism but lost, facing heavy legal costs.
- The Rennes-le-Château conspiracy culture illustrates hyperreality, blurring lines between fact and fiction in media landscapes.
Sidephone: A minimalist Android phone with swappable USB keypads
7 hours ago
- Sidephone supports making and receiving calls, and has a dedicated Contacts app.
- It features SMS messaging without RCS, a retro-style Camera app, and a Clock app for timers and alarms.
- Device includes Files and Gallery apps, Settings for customization, and a custom Maps app by HERE.
- LocalSend for cross-platform file sharing, Voice Recorder, and Auxio as a local music player.
- Provides secure messaging, Proton suite apps (Mail, VPN, Calendar, Drive, Pass), and Standard Notes for secure note-taking.
- Offers browser choices, email client via K-9, cloud sync with Davx5, and apps like AntennaPod and BreezyWeather.
- Includes basic apps (Calculator, Signal, Uber), Obtanium for app updates, and compatibility for music streaming and third-party messaging.
- Authentication apps support (DUO, Aegis, BitWarden) and offers like free shipping over $299, one-year warranty, and 30-day return policy.
4-bit floating point FP4
6 hours ago
- Historically, floating point numbers evolved from 32-bit to 64-bit precision, with C using 'float' for 32-bit and 'double' for 64-bit, while Python uses 'float' for 64-bit.
- Neural networks drive demand for lower precision floating point formats like 16-bit, 8-bit (FP8), and 4-bit (FP4) to fit more parameters in memory.
- FP4 uses one bit for sign and three bits for exponent and mantissa, with common formats including E3M0, E2M1, E1M2, and E0M3, where E2M1 is widely supported.
- The value of an FP4 number is calculated as (−1)^s * 2^(e−b) * (1 + m/2), with a bias (b) to allow positive and negative exponents, except when e=0.
- A 16-value table for E2M1 format shows representable numbers from -6 to +6, including ±0, illustrating the distribution on log and linear scales.
- The Pychop library emulates reduced-precision formats like FP4, with code provided to generate and display the E2M1 table of values.
- Future discussions may cover other 4-bit formats like NF4, which better match the weight distributions in large language models (LLMs).
Floating Point Fun on Cortex-M Processors
2 days ago
- The article explains Arm's floating point ABIs: soft, softfp, and hard, and the resulting linker errors when mixing them.
- Soft ABI emulates floating-point operations in software, softfp uses FPU instructions but passes arguments in integer registers, and hard uses FPU instructions and registers.
- The nRF52840 and ESP32-S3 examples illustrate how different ABIs affect code generation, with specific assembly outputs shown.
- Zephyr's configuration options (CONFIG_FPU, CONFIG_FP_HARDABI, CONFIG_FP_SOFTABI) control the FPU and ABI selection.
- Enabling the FPU dynamically is possible but risky; an example demonstrates handling No Coprocessor (NOCP) faults.
- Floating point operations without FPU enabled cause exceptions, but can be managed via runtime enabling or fault handlers.
- The article hints at future discussions on trade-offs between hardware and software floating point implementations.
Show HN: AI Subroutines – Run automation scripts inside your browser tab
a day ago
- Web agents often fail at scaling due to token costs, latency, and non-determinism, with direct API calls hindered by complex authentication mechanisms.
- rtrvr solves this by recording and replaying tasks within the user's browser, using the page's own context for auth, CSRF, and fingerprints, avoiding out-of-process rebuilding.
- Network captures are ranked and trimmed to identify relevant API calls, filtering out noise like analytics beacons, and avoiding volatile operation identifiers that break on site updates.
- Subroutines are deterministic, zero-token tools that agents call with parameters, running as scripts in-page for tasks like sending DMs or posting on social media.
- The rtrvr.* helpers provide functions for DOM interactions and authenticated requests, enabling reliable automation without brittle selectors or manual fetch handling.
- The approach differs from other tools by keeping execution in-page for native auth handling, aiming to build a community library of Subroutines for common web actions.
- Additional features include BYO LLM subscriptions, WhatsApp control, Knowledge Base upgrades, and site analytics via Rover, with the extension available for testing and feedback.
College instructor turns to typewriters to curb AI-written work
4 hours ago
- A German instructor at Cornell University, Grit Matthias Phelps, uses manual typewriters in class to counter students' reliance on AI and online tools for writing assignments.
- This 'analog' assignment aims to recreate a pre-digital classroom experience, emphasizing focused, slow-paced work without digital distractions or easy corrections.
- Students found the typewriters challenging but educational, noting benefits like fewer distractions, more socialization, and having to think independently without AI assistance.
- The exercise is part of a broader trend in education to adopt old-school testing methods, such as in-person exams, to prevent AI misuse in academic work.
- Despite initial difficulties, students appreciated the unique learning experience, with some embracing the physical and creative aspects of typewriter use.
Thoughts and feelings around Claude Design
4 hours ago
- The article critiques how Figma became dominant through systematization with its own primitives, but these are poorly documented and not part of LLM training data, making it less relevant in the agentic era.
- As AI makes code easier for designers and source of truth shifts back to code, Figma's complex infrastructure may become obsolete, and manually updating design changes between code and Figma is tedious.
- Claude Design, in contrast, embraces 'truth to materials' by being based on HTML and JS, with potential seamless integration with Claude Code, reducing friction between design and implementation.
- The future of design tools may fork: one like Claude Design that integrates with code, and another that is a pure exploration environment without code constraints, possibly for sketching or high-fidelity compositing.
- The article warns that Figma could face a decline similar to Sketch's if it doesn't adapt, urging Figma to innovate and Sketch to add more creative effects instead of relying on past success.
Traders placed over $1B in perfectly timed bets on the Iran war
5 hours ago
- Online betting markets like Polymarket and Kalshi enabled large, well-timed bets predicting major events in the US-Israel war with Iran, raising insider trading concerns.
- Significant bets included $100,000 wins on US airstrikes, over $550,000 on Ayatollah Khamenei's removal, and $950 million on oil price drops before ceasefire announcements.
- Suspicious trading patterns, such as surges in oil futures and prediction market bets just before major announcements, prompted investigations by the CFTC and lawmakers.
- Legal challenges arise as insider trading laws for commodity futures and prediction markets are underdeveloped, and anonymous blockchain trading complicates enforcement.
- Regulatory efforts face hurdles due to jurisdictional disputes between federal agencies like the CFTC and state legislatures, and a lack of full commission appointments.
- Experts warn that insider trading on geopolitical events can distort markets and erode trust in governments, with calls for stricter regulations and enforcement.
Show HN: Remoroo. trying to fix memory in long-running coding agents
4 days ago
- Remoroo autonomously conducts research on code overnight, performing edits, tests, evaluations, and reverts to improve results.
- It contrasts manual ML research, which is slow and often unproductive, with Remoroo's efficient, automated experimentation that yields verified, reproducible outcomes.
- Users specify a research goal in a file like program.md, and Remoroo plans, edits, trains, and evaluates experiments based on a fixed metric, such as val_bpb.
- The tool runs multiple experiments, keeping successful ones and discarding failures, with examples showing reductions in validation loss through architecture and schedule searches.
- Remoroo is billed in credits based on runtime, offers a free tier, and emphasizes verified results over guessing, distinguishing it from typical coding agents.
Georgia's Voting Technology Blunder
7 hours ago
- History of voting machine flaws post-Bush v. Gore, including failed standardization attempts and censorship via DMCA.
- Critique of current voting technology, especially Georgia's plan to use OCR for ballot verification, deemed insecure.
- Impact of partisan politics on voting machine criticism, leading to schismogenic reversal where valid concerns are dismissed.
- Advocacy for simpler, more secure voting methods like hand-marked paper ballots counted by existing scanners.
- Miscellaneous links and updates on tech policy, appearances, and upcoming books.
Building a Grow-Only Counter on a Sequentially Consistent KV Store
5 days ago
- The post discusses building a Grow-Only Counter (G-Counter) on Maelstrom's sequentially consistent key-value store (SeqKV), addressing Challenge #4 in Fly.io's distributed systems series.
- Initial naive solutions using a single key or per-node keys with read-modify-write or CompareAndSwap (CAS) operations are non-deterministic due to concurrent updates and stale reads allowed under sequential consistency, where anomalies like stale reads after a cooldown period can occur.
- A CRDT-like approach is explored, where each node updates its own counter key, reducing write conflicts by summing all node counters for reads, but this still faces non-determinism due to SeqKV's consistency model and implementation quirks.
- The post explains sequential consistency as a model where operations have a total order consistent with each client's program order but may not match real-time order, allowing anomalies like stale reads, which differs from linearizability that enforces real-time order.
- A hacky solution involves writing a unique value to force SeqKV clients to the latest state before reads, leveraging SeqKV's internal behavior, while non-hacky alternatives include implementing a true CRDT with gossip or switching to a linearizable store like LinKV for deterministic results.

first prev0next

About|Login

Bilingual

Adobe Has Run Out of Allies
4 hours ago
- Adobe shifted from Creative Suite to Creative Cloud in 2013, promising more agility, updates, and affordability.
- Initially, Adobe's developers and product managers actively engaged with media and users, showing care and pride.
- Adobe's focus shifted from individual creative users to corporate clients, altering its media strategy and communication.
- The company replaced customer-facing teams with business-oriented managers and legal messaging, alienating artists.
- Adobe now faces competition from Apple, Blackmagic, and Canva, with declining stock value and user loyalty.
- Individual users and shareholders have turned against Adobe, leaving the company isolated and widely detested.
Hardware Is Hard?
4 hours ago
- The author, feeling bored with AI taking over their work, decides to build a robot for fun, inspired by nostalgic college projects.
- Researching robot building reveals a lack of electronics knowledge, including basics like electricity, components, and tools such as multimeters and breadboards.
- To start simply, the author chooses to build a "Hello, World!" equivalent in robotics: a small remote-controlled robot car using a phone.
- The core principle is that voltage difference moves motors, explained through Ohm's law (V=IR) with an analogy to water flowing downhill.
- Required components include a 4WD chassis kit, 18650 Li-ion batteries, ESP32 as the brain, TB6612FNG motor controller, multimeter, and other tools like a soldering iron.
- Assembly involves testing motor directions, wiring components, and using the ESP32 to control motors via GPIO pins and pulse width modulation (PWM) for speed.
- The TB6612FNG module acts as an H-bridge to control motor direction and speed based on ESP32 signals, avoiding energy waste from resistance.
- Programming the ESP32 involves C code interacting with memory-mapped I/O to trigger pins, with help from online resources and AI like Claude.
- Wiring follows detailed instructions, and despite challenges like faulty components, the author succeeds in making the wheels move, achieving a eureka moment.
- The author concludes that electronics revolves around voltage, appreciates AI assistance, and encourages embracing life's richness through hands-on projects.
PostgreSQL production incident caused by transaction ID wraparound
3 hours ago
- PostgreSQL transaction ID wraparound is a silent but severe failure mode that can cause a complete write outage.
- It occurs when freezing of old row versions doesn't happen, allowing transaction IDs to approach a hard safety limit.
- The failure is not immediate; it can manifest after months or years of normal operation with no warning signs.
- In the described incident, disabling autovacuum on tables, including unused ones, led to transaction ID aging.
- Once the limit is reached, PostgreSQL intentionally blocks all writes to prevent data corruption, making the database read-only.
- Recovery requires manual VACUUM FREEZE operations to advance transaction IDs beyond the danger zone.
- Transaction ID consumption depends on write rate and time; even modest workloads can lead to risk after about 7.5 months.
- Unlike PostgreSQL, SQL Server avoids this issue by using Log Sequence Numbers (LSNs) instead of finite, reusable transaction IDs.
- Understanding transaction ID math and treating autovacuum as a safety mechanism is crucial for PostgreSQL production stability.
Optimizing Ruby Path Methods
3 hours ago
- Bootsnap optimizes Ruby's file loading by caching load paths to avoid repeated expensive searches.
- Load path caching converts O(N*M) boot time complexity into O(1) lookups, speeding up application startup.
- Cache invalidation relies on directory mtime, but git doesn't update mtimes, requiring frequent cache rebuilds on CI.
- Bootsnap's path scanner had an N+1 syscall issue due to stat calls for each directory entry.
- Dir.scan was introduced in Ruby 4.1 to expose file types without extra stat calls, improving scanning speed by 2x.
- File.join was optimized by implementing a fast path for ASCII-compatible encodings and fixing inefficient reverse search.
- Encoding handling, especially for Shift JIS, previously slowed File.join; a fast path now speeds up UTF-8 and ASCII cases.
- Optimizations to File.join made it over 7 times faster for common cases, outperforming simple string interpolation.
- Other path methods like File.basename, File.extname, and File.dirname received similar optimizations for consistency.
I learned Unity the wrong way
3 days ago
- The author shares their experience of learning Unity in an inefficient or incorrect manner
- They reflect on common pitfalls and mistakes made during the learning process
- The article likely provides advice on better approaches to learning Unity effectively
- It emphasizes human-written content focused on personal development insights
Shader Lab, like Photoshop but for shaders
2 days ago
- The text appears to be a UI settings panel for a CRT (Cathode Ray Tube) effect, likely from a video editing or graphics software.
- It lists various properties and adjustment sliders under categories such as General, Settings, Distortion, Phosphor, Persistence, Noise, Glitch, and Bloom.
- Key adjustable parameters include Opacity, Blend Mode, Hue, Saturation, Mask Scale, Scanline Intensity, Mask Intensity, Barrel Distortion, Convergence, Vignette, Beam Focus, Brightness, Highlight Drive, Shoulder, Chroma Retention, Shadow Lift, Flicker, Glitch Speed, and Bloom settings (Intensity, Threshold, Radius, Softness).
IPC medley: message-queue peeking, io_uring, and bus1
2 days ago
- Proposal for mq_timedreceive2() adds flags and index parameters for peeking at message queues, useful for monitoring and checkpoint/restore.
- io_uring IPC patch series aims to create a high-bandwidth communication mechanism using shared ring buffers, similar to D-Bus, but needs more development.
- Bus1 returns rewritten in Rust, offering kernel-mediated IPC with message and capability passing, seeking integration help from Rust for Linux community.
Year of the IPv6 Overlay Network
a day ago
- Nebula 1.10 introduces IPv6 support in overlay networks, enabling future-proofing and seamless connectivity.
- A new v2 certificate format using ASN.1 replaces Protobufs, supporting IPv6 and multiple IPs per host.
- Multiple Nebula IPs per host can now be assigned, including both IPv4 and IPv6 addresses.
- IPv6 adoption eliminates addressing conflicts and prepares networks for IPv6 dominance.
- Upgrade process involves upgrading to v1.10, creating a v2 CA, and minting v2 certs with IPv4/IPv6 addresses.
- Firewall behavior changes: default_local_cidr_any now defaults to false for improved security.
- PKCS#11 support for P256 keys enables enterprise HSM and smart card integration.
- ECMP support for unsafe_routes provides load balancing and redundancy with multiple gateways.
- SO_MARK socket option support on Linux allows advanced routing via firewall marks.
- Enhanced library integration offers configurable logging and versioning for embedded use.
Comparing GPT-5.4, Opus 4.6, GLM-5.1, Kimi K2.5, MiMo V2 Pro and MiniMax M2.7
6 hours ago
- A benchmarking experiment compared six large language models (GPT-5.4, Opus 4.6, GLM-5.1, Kimi K2.5, MiMo V2 Pro, MiniMax M2.7) for building a native macOS app to manage drive ejection issues.
- All models performed similarly in planning and compilation phases, with MiMo V2 Pro notably using kernel API instead of shell commands, though this was considered irrelevant.
- Post-compilation, some apps crashed initially, but all could be fixed with comparable effort, highlighting a lack of end-to-end testing tools for native apps as a limitation.
- In code quality, GPT-5.4 and GLM-5.1 were preferred for leaner, more maintainable output, though prompting could guide any model toward cleaner code.
- A peer review ranking placed GPT-5.4 and Opus 4.6 consistently in the top three, with GPT-5.4 leading; MiMo V2 Pro also ranked high but scored lower on cleanliness, while GLM-5.1 excelled in cleanliness but placed fourth overall.
- Kimi K2.5 and MiniMax M2.7 consistently ranked at the bottom, correlating with their lower cost, and models showed little self-bias in scoring, except GPT-5.4, which ranked itself first, aligning with consensus.
- The conclusion recommends any top model for practical use, noting GLM and MiMo offer competitive performance at lower costs, and emphasizes picking one model to avoid procrastination.
The Mystery of Rennes-Le-Château, Part 4: Non-Fiction Meets Fiction
10 hours ago
- The Holy Blood and the Holy Grail was published in 1982, gaining strong sales and cultural traction on both sides of the Atlantic.
- Its success was influenced by timing, with Raiders of the Lost Ark's popularity and Britain's puzzle-book craze post-Masquerade.
- The U.S. paperback edition had an initial print run of 500,000 copies; translations and millions sold made the authors wealthy.
- Critics dismissed the book as shoddy history, but literary figures like Anthony Burgess praised its novelistic approach.
- Pierre Plantard distanced himself from the book's claim that Merovingian kings were descendants of Jesus Christ.
- Subsequent works like The Messianic Legacy (1987) saw declining impact, with Henry Lincoln obsessively revisiting theories.
- Graham Hancock emerged as a prominent figure in pseudo-history with books like Fingerprints of the Gods.
- Rennes-le-Château lore influenced media, including adventure games like Broken Sword and Gabriel Knight 3.
- The Tomb of God (1996) by Schellenberger and Andrews revived interest, proposing Jesus's body was hidden near Rennes-le-Château.
- Skeptics like Umberto Eco (Foucault's Pendulum) and BBC's Timewatch debunked many conspiracy claims.
- Dan Brown's The Da Vinci Code (2003) became a global phenomenon, borrowing heavily from the conspiracy theories.
- Leigh and Baigent sued Brown for plagiarism but lost, facing heavy legal costs.
- The Rennes-le-Château conspiracy culture illustrates hyperreality, blurring lines between fact and fiction in media landscapes.
Sidephone: A minimalist Android phone with swappable USB keypads
7 hours ago
- Sidephone supports making and receiving calls, and has a dedicated Contacts app.
- It features SMS messaging without RCS, a retro-style Camera app, and a Clock app for timers and alarms.
- Device includes Files and Gallery apps, Settings for customization, and a custom Maps app by HERE.
- LocalSend for cross-platform file sharing, Voice Recorder, and Auxio as a local music player.
- Provides secure messaging, Proton suite apps (Mail, VPN, Calendar, Drive, Pass), and Standard Notes for secure note-taking.
- Offers browser choices, email client via K-9, cloud sync with Davx5, and apps like AntennaPod and BreezyWeather.
- Includes basic apps (Calculator, Signal, Uber), Obtanium for app updates, and compatibility for music streaming and third-party messaging.
- Authentication apps support (DUO, Aegis, BitWarden) and offers like free shipping over $299, one-year warranty, and 30-day return policy.
4-bit floating point FP4
6 hours ago
- Historically, floating point numbers evolved from 32-bit to 64-bit precision, with C using 'float' for 32-bit and 'double' for 64-bit, while Python uses 'float' for 64-bit.
- Neural networks drive demand for lower precision floating point formats like 16-bit, 8-bit (FP8), and 4-bit (FP4) to fit more parameters in memory.
- FP4 uses one bit for sign and three bits for exponent and mantissa, with common formats including E3M0, E2M1, E1M2, and E0M3, where E2M1 is widely supported.
- The value of an FP4 number is calculated as (−1)^s * 2^(e−b) * (1 + m/2), with a bias (b) to allow positive and negative exponents, except when e=0.
- A 16-value table for E2M1 format shows representable numbers from -6 to +6, including ±0, illustrating the distribution on log and linear scales.
- The Pychop library emulates reduced-precision formats like FP4, with code provided to generate and display the E2M1 table of values.
- Future discussions may cover other 4-bit formats like NF4, which better match the weight distributions in large language models (LLMs).
Floating Point Fun on Cortex-M Processors
2 days ago
- The article explains Arm's floating point ABIs: soft, softfp, and hard, and the resulting linker errors when mixing them.
- Soft ABI emulates floating-point operations in software, softfp uses FPU instructions but passes arguments in integer registers, and hard uses FPU instructions and registers.
- The nRF52840 and ESP32-S3 examples illustrate how different ABIs affect code generation, with specific assembly outputs shown.
- Zephyr's configuration options (CONFIG_FPU, CONFIG_FP_HARDABI, CONFIG_FP_SOFTABI) control the FPU and ABI selection.
- Enabling the FPU dynamically is possible but risky; an example demonstrates handling No Coprocessor (NOCP) faults.
- Floating point operations without FPU enabled cause exceptions, but can be managed via runtime enabling or fault handlers.
- The article hints at future discussions on trade-offs between hardware and software floating point implementations.
Show HN: AI Subroutines – Run automation scripts inside your browser tab
a day ago
- Web agents often fail at scaling due to token costs, latency, and non-determinism, with direct API calls hindered by complex authentication mechanisms.
- rtrvr solves this by recording and replaying tasks within the user's browser, using the page's own context for auth, CSRF, and fingerprints, avoiding out-of-process rebuilding.
- Network captures are ranked and trimmed to identify relevant API calls, filtering out noise like analytics beacons, and avoiding volatile operation identifiers that break on site updates.
- Subroutines are deterministic, zero-token tools that agents call with parameters, running as scripts in-page for tasks like sending DMs or posting on social media.
- The rtrvr.* helpers provide functions for DOM interactions and authenticated requests, enabling reliable automation without brittle selectors or manual fetch handling.
- The approach differs from other tools by keeping execution in-page for native auth handling, aiming to build a community library of Subroutines for common web actions.
- Additional features include BYO LLM subscriptions, WhatsApp control, Knowledge Base upgrades, and site analytics via Rover, with the extension available for testing and feedback.
College instructor turns to typewriters to curb AI-written work
4 hours ago
- A German instructor at Cornell University, Grit Matthias Phelps, uses manual typewriters in class to counter students' reliance on AI and online tools for writing assignments.
- This 'analog' assignment aims to recreate a pre-digital classroom experience, emphasizing focused, slow-paced work without digital distractions or easy corrections.
- Students found the typewriters challenging but educational, noting benefits like fewer distractions, more socialization, and having to think independently without AI assistance.
- The exercise is part of a broader trend in education to adopt old-school testing methods, such as in-person exams, to prevent AI misuse in academic work.
- Despite initial difficulties, students appreciated the unique learning experience, with some embracing the physical and creative aspects of typewriter use.
Thoughts and feelings around Claude Design
4 hours ago
- The article critiques how Figma became dominant through systematization with its own primitives, but these are poorly documented and not part of LLM training data, making it less relevant in the agentic era.
- As AI makes code easier for designers and source of truth shifts back to code, Figma's complex infrastructure may become obsolete, and manually updating design changes between code and Figma is tedious.
- Claude Design, in contrast, embraces 'truth to materials' by being based on HTML and JS, with potential seamless integration with Claude Code, reducing friction between design and implementation.
- The future of design tools may fork: one like Claude Design that integrates with code, and another that is a pure exploration environment without code constraints, possibly for sketching or high-fidelity compositing.
- The article warns that Figma could face a decline similar to Sketch's if it doesn't adapt, urging Figma to innovate and Sketch to add more creative effects instead of relying on past success.
Traders placed over $1B in perfectly timed bets on the Iran war
5 hours ago
- Online betting markets like Polymarket and Kalshi enabled large, well-timed bets predicting major events in the US-Israel war with Iran, raising insider trading concerns.
- Significant bets included $100,000 wins on US airstrikes, over $550,000 on Ayatollah Khamenei's removal, and $950 million on oil price drops before ceasefire announcements.
- Suspicious trading patterns, such as surges in oil futures and prediction market bets just before major announcements, prompted investigations by the CFTC and lawmakers.
- Legal challenges arise as insider trading laws for commodity futures and prediction markets are underdeveloped, and anonymous blockchain trading complicates enforcement.
- Regulatory efforts face hurdles due to jurisdictional disputes between federal agencies like the CFTC and state legislatures, and a lack of full commission appointments.
- Experts warn that insider trading on geopolitical events can distort markets and erode trust in governments, with calls for stricter regulations and enforcement.
Show HN: Remoroo. trying to fix memory in long-running coding agents
4 days ago
- Remoroo autonomously conducts research on code overnight, performing edits, tests, evaluations, and reverts to improve results.
- It contrasts manual ML research, which is slow and often unproductive, with Remoroo's efficient, automated experimentation that yields verified, reproducible outcomes.
- Users specify a research goal in a file like program.md, and Remoroo plans, edits, trains, and evaluates experiments based on a fixed metric, such as val_bpb.
- The tool runs multiple experiments, keeping successful ones and discarding failures, with examples showing reductions in validation loss through architecture and schedule searches.
- Remoroo is billed in credits based on runtime, offers a free tier, and emphasizes verified results over guessing, distinguishing it from typical coding agents.
Georgia's Voting Technology Blunder
7 hours ago
- History of voting machine flaws post-Bush v. Gore, including failed standardization attempts and censorship via DMCA.
- Critique of current voting technology, especially Georgia's plan to use OCR for ballot verification, deemed insecure.
- Impact of partisan politics on voting machine criticism, leading to schismogenic reversal where valid concerns are dismissed.
- Advocacy for simpler, more secure voting methods like hand-marked paper ballots counted by existing scanners.
- Miscellaneous links and updates on tech policy, appearances, and upcoming books.
Building a Grow-Only Counter on a Sequentially Consistent KV Store
5 days ago
- The post discusses building a Grow-Only Counter (G-Counter) on Maelstrom's sequentially consistent key-value store (SeqKV), addressing Challenge #4 in Fly.io's distributed systems series.
- Initial naive solutions using a single key or per-node keys with read-modify-write or CompareAndSwap (CAS) operations are non-deterministic due to concurrent updates and stale reads allowed under sequential consistency, where anomalies like stale reads after a cooldown period can occur.
- A CRDT-like approach is explored, where each node updates its own counter key, reducing write conflicts by summing all node counters for reads, but this still faces non-determinism due to SeqKV's consistency model and implementation quirks.
- The post explains sequential consistency as a model where operations have a total order consistent with each client's program order but may not match real-time order, allowing anomalies like stale reads, which differs from linearizability that enforces real-time order.
- A hacky solution involves writing a unique value to force SeqKV clients to the latest state before reads, leveraging SeqKV's internal behavior, while non-hacky alternatives include implementing a true CRDT with gossip or switching to a linearizable store like LinKV for deterministic results.

Hasty Briefsbeta

Adobe Has Run Out of Allies

Hardware Is Hard?

PostgreSQL production incident caused by transaction ID wraparound

Optimizing Ruby Path Methods

I learned Unity the wrong way

Shader Lab, like Photoshop but for shaders

IPC medley: message-queue peeking, io_uring, and bus1

Year of the IPv6 Overlay Network

Comparing GPT-5.4, Opus 4.6, GLM-5.1, Kimi K2.5, MiMo V2 Pro and MiniMax M2.7

The Mystery of Rennes-Le-Château, Part 4: Non-Fiction Meets Fiction

Sidephone: A minimalist Android phone with swappable USB keypads

4-bit floating point FP4

Floating Point Fun on Cortex-M Processors

Show HN: AI Subroutines – Run automation scripts inside your browser tab

College instructor turns to typewriters to curb AI-written work

Thoughts and feelings around Claude Design

Traders placed over $1B in perfectly timed bets on the Iran war

Show HN: Remoroo. trying to fix memory in long-running coding agents

Georgia's Voting Technology Blunder

Building a Grow-Only Counter on a Sequentially Consistent KV Store

Adobe Has Run Out of Allies

Hardware Is Hard?

PostgreSQL production incident caused by transaction ID wraparound

Optimizing Ruby Path Methods

I learned Unity the wrong way

Shader Lab, like Photoshop but for shaders

IPC medley: message-queue peeking, io_uring, and bus1

Year of the IPv6 Overlay Network

Comparing GPT-5.4, Opus 4.6, GLM-5.1, Kimi K2.5, MiMo V2 Pro and MiniMax M2.7

The Mystery of Rennes-Le-Château, Part 4: Non-Fiction Meets Fiction

Sidephone: A minimalist Android phone with swappable USB keypads

4-bit floating point FP4

Floating Point Fun on Cortex-M Processors

Show HN: AI Subroutines – Run automation scripts inside your browser tab

College instructor turns to typewriters to curb AI-written work

Thoughts and feelings around Claude Design

Traders placed over $1B in perfectly timed bets on the Iran war

Show HN: Remoroo. trying to fix memory in long-running coding agents

Georgia's Voting Technology Blunder

Building a Grow-Only Counter on a Sequentially Consistent KV Store