Hasty Briefsbeta

Bilingual

The LLMs Yearn for the Spines
a day ago
- LLMs increasingly use the word 'spine' in code project PRs, with a 20x increase in 2026 compared to 2025.
- The spike in 'spine' usage is not explained by overall GitHub growth, as overall PRs only increased 1.5x in the same period.
- The word 'spine' is not specific to one LLM; it appears across models like Claude and Cursor, with potential contributions from GPT 5.5 and Opus 4.8.
- Other LLM-favored words include 'gates' (1 in 300 PRs in titles), 'lane', 'proofs', 'truth', and 'seam'.
- 'Ladder' is a false positive in PR bodies due to automated boto3 upgrades, but looking at titles only removes this artifact.
- The author developed a tool to analyze PR data and shared it as a gist, inviting feedback on methodology.
- The analysis suggests LLMs have stylistic preferences for certain words, potentially useful for identifying AI-generated code.
- Formal verification and proofs are also trending topics in PRs, reflecting AI's influence.
- The author notes limitations, such as not filtering code samples by date and relying on PR titles as a proxy.
Why MCP Was Always a Bad Idea
an hour ago
- MCP was introduced in November 2024 as a protocol for connecting LLMs to external services, but models have since become far more capable.
- Context bloat from multiple MCP servers led to workarounds like generic search/execute patterns, but these are temporary fixes.
- Modern LLMs can write scripts, call APIs directly, and use CLI help commands, making most MCP servers redundant since they wrap existing APIs.
- Proposed solution: standardize direct HTTP API usage with headers like Accept: text/markdown and Accept-Language for agent-friendly responses.
- The article argues that MCP is now obsolete and should be replaced by direct HTTP APIs and CLIs.
What's been going on in w64devkit the past year
6 hours ago
- Release security improved: code-signed binaries, automated GitHub Actions builds, and immutable release artifacts.
- x64 release now a multilib toolchain, able to compile 32-bit Windows programs via -m32 or i686-w64-mingw32 prefix.
- Compiler no longer requires bin/ on PATH; GCC patched for direct invocation.
- Bigobj COFF format now automatically used when needed, thanks to upstream Binutils patch by Peter0x44.
- Builds switched from -Os to -O2 for static runtimes and computationally expensive tools, increasing size and speed.
- CMake and Ninja included, with CMake defaulting to Ninja; also added ccmake TUI and Ninja Multi-Config support.
- Ccache included with Windows-specific patches; can be enabled via w64devkit.ini.
- New tools: widl/uuidgen for COM, make2compdb for JSON compilation databases, recycle for sending files to recycle bin, quilt for patch management, NSIS for installers, zstd for compression.
- C11 threads implementation added to Mingw-w64 runtime, requiring Windows 7+, excluding recursive locks.
- std::terminate now traps instead of calling exit, reducing binary size and enabling GDB inspection.
- Future plans include re-enabling LTO with FatLTO runtimes for C++ and Fortran, using AI to fix LTO bugs.
- Commitment to Windows XP support for x86 release maintained.
Frontier Labs Are Selling Garbage to Fools in Washington
an hour ago
- AI executives exaggerate existential risks to gain regulatory favors and protect profits, turning congressional hearings into theatrical performances.
- Supposed AI security breaches were caused by basic incompetence, such as misconfigured firewalls and carelessly exposed credentials, not superintelligent rogue agents.
- Andrew Yang's testimony about self-replicating alien code was a misinterpretation of routine synthetic data generation, not a digital pandemic.
- The real motive behind calls for a slowdown is to form a government-sanctioned cartel that locks in current market leaders and suppresses open-source competition.
- Open-weight models threaten proprietary profits, so frontier labs seek regulatory moats to criminalize independent development and maintain monopoly pricing.
Software Sandboxing: The Basics
2 hours ago
- Sandboxing is defined as discretionary privilege dropping: programmatically restricting privileges without administrative authority.
- Programmatic privilege dropping requires OS-specific interfaces like FreeBSD's Capsicum and Linux's Seccomp, as traditional UNIX permissions are inadequate.
- Avoid relying on superuser mechanisms (e.g., suid binaries) for sandboxing; Linux namespaces introduce security risks and are better suited for containerization.
- The actor model (processes as actors) combined with capability-based security (using file descriptors as capabilities) provides a robust framework for compartmentalized sandboxing.
- Capsicum (FreeBSD) is highlighted as a superior sandboxing mechanism, requiring only ~100 lines of code to integrate, compared to thousands for other systems.
- Non-blocking I/O on UNIX presents challenges; careful handling of file descriptors and use of proactors or io_uring is needed to avoid DoS vulnerabilities.
- Existing code can be sandboxed via libc interposition (e.g., LD_PRELOAD) using tools like Super Capsicumizer 9000 or Emilua's libc_service module.
- Seccomp (Linux) is complex and error-prone due to multiarch issues and syscall numbering; Kafel helps define syscall policies but has limitations.
- Practical examples with Emilua show how to spawn sandboxed actors, restrict file access, and handle legacy code through dynamic policy enforcement.
- Threat models must be evaluated critically: sandboxing is not always necessary but crucial for software handling sensitive data (e.g., Telegram's tdlib).
Self-hosted inference orchestrators compared: LocalAI, exo, GPUStack, vLLM
3 hours ago
- Survey compares self-hosted orchestrators for OpenAI-compatible endpoints on GPU machines, covering 11 tools as of September 2026.
- Ollama is the default for single-machine use; llama.cpp and vLLM are engines that others wrap; LiteLLM is a router not a runtime.
- LocalAI offers broad modality support with distributed mode via P2P and NATS; exo excels on Apple Silicon with zero-config multi-machine.
- GPUStack and Xinference provide enterprise consoles with users, keys, and dashboards; NVIDIA Dynamo targets datacenter-scale disaggregated inference.
- SkyPilot and dstack burst jobs to clouds but are not inference servers; CoderAI adds escalation, multi-modality, and training across nodes.
- Recommendations: pick Ollama for simplicity, exo for Macs, LocalAI for breadth, GPUStack/Xinference for departments, and CoderAI for escalation and full modality support.
The shape that does not return
a day ago
- A system is the product of interactions between parts, not the sum of the parts themselves.
- Optimizing individual parts in isolation often degrades overall system performance, especially in tightly coupled systems.
- Reversibility is a key criterion for safe optimization: can you undo a change without catastrophic loss?
- Optimization reduces internal variety, which under Ashby's law reduces adaptive capacity to absorb disturbances.
- Irreversible optimization can produce fragility by consuming the ability to course-correct and learn.
- Stafford Beer's 'relaxation' refers to preserving un-collapsed variety as internal slack for viability.
- Hysteresis shows that reversal may not restore original state due to path-dependent changes in internal structure.
- Reversibility is observer-relative: what is reversible for one stakeholder may be irreversible for another (e.g., factory closure for investor vs. community).
- The boundary of optimization determines who bears costs and who experiences lock-in; it is a claim about the relevant whole.
- Von Foerster's ethical imperative: act to increase the number of choices; irreversible optimization violates this.
- Zero-harm optimization may be impossible; the focus should shift to minimizing the worst possible outcome (minimax principle).
Show HN: Three genlocked RP2350B make a console – 3k sprite pixels per line)
3 hours ago
- PapyDeck is a Lua-based game console with a sandboxed VM that prints line numbers on errors instead of crashing.
- Games are plain folders on a micro-SD card containing a main.lua and art; no compiled blobs or locked cartridges.
- The same VM runs in a browser simulator bit-for-bit, allowing code editing and instant replay.
- Hardware uses three genlocked RP2350B chips (six cores) for HDMI output with 3,000 opaque sprite pixels per scanline and no tearing.
- Current V1 alpha supports HDMI, sprites, tile backgrounds, micro-SD card loading, Bluetooth gamepads, Wi-Fi, audio, and OTA firmware updates.
- Hardware is under CERN-OHL-P, firmware under MIT; repositories will be open for forking and manufacturing.
- The console can list and download published games over Wi-Fi directly onto the card.
- Still open items include HDMI audio, verified leaderboard, and final PCB with voltage supervisor and secure element.
As White House shields the AI gold rush, Trump family and other allies strike...
4 hours ago
- Trump's sons secured $620m Pentagon loan, Marine Corps robotics contract, and undisclosed Air Force drone deal through AI-linked ventures in the past year.
- Longtime Trump friend Michael Dell won nearly $9bn Pentagon contract.
- Trump aggressively promotes AI on Truth Social, opposing guardrails and claiming only a 'strong president' is needed, amid sinking public support.
- Conflict of interest concerns arise as Trump's family and allies accumulate financial stakes in AI infrastructure and defense tech, while Trump resists new regulations.
- Administration cites competitiveness with China as reason for resisting rules; David Sacks dismisses AI risks as 'political theater'.
- Industry insiders resign with warnings: AI may 'kill us all by the end of the decade', and an AI-built worm could compromise billions of devices.
- Polling shows 11% support local AI datacenters, 57% disapprove of Trump's AI handling; House votes 417-3 to make datacenters pay for grid upgrades.
- Trump brothers' investments include Dominari Holdings, American Data Centers, Vulcan Elements (Pentagon loan), Foundation Future Industries (Marine Corps deal), and Powerus (Air Force contract).
- Sacks lobbied Trump to scrap AI executive order; his firm holds stakes in SpaceX and AI startups. Musk, Zuckerberg, Huang also lobbied against AI oversight body.
- Ethics watchdogs question Pentagon contracts to Trump-connected firms; Democratic lawmakers seek investigation.
- Tech industry donations still favor Democrats, but some AI leaders like Altman and Brockman donated to Trump's campaign.
- Trump to host leading AI companies and Chinese president Xi Jinping at White House next week.
US Revokes Limits on Power Plants' Climate Pollution
4 hours ago
- The EPA repealed the 2024 Carbon Pollution Standards that required coal and gas plants to capture 90% of carbon emissions or shut down by 2039.
- The repeal eliminates most limits on power plant carbon emissions, representing a major attack on climate action.
- The 2024 standards would have reduced carbon pollution by 1.38 billion metric tons by 2047 and prevented 1,200 deaths and 360,000 asthma attacks in 2035 alone.
- Human Rights Watch documented how coal plant pollution harms nearby communities, citing examples in Bulgaria, Bosnia and Herzegovina, and Türkiye.
- The EPA claimed $370 million in regulatory cost savings but ignored public health costs, and earlier stopped factoring health costs into economic impact estimates.
- The EPA also proposed removing all greenhouse gas requirements for power plants and revoked the 2009 endangerment finding linking emissions to climate change.
- The article calls for restoring the 2024 standards and strengthening fossil fuel regulation.
Trying the Software Factory Pattern
4 hours ago
- The AI ecosystem in 2026 evolves rapidly, requiring continuous adoption of new patterns like Claude Code, workspace reorganization, and migration from Jira to Linear.
- The software factory pattern involves looping on a broad goal with an orchestrated harness driving progress, enabling autonomous task management.
- Implementation uses an agent skill /linear-project-loop that audits project goals, metrics, and issues, then autonomously works on tasks such as writing PRs or updating issues.
- The pattern forces recognition of hidden state (e.g., goals) and enables post-release monitoring of metrics like adoption spikes or error rates.
- Success of the factory pattern depends on having all underlying tools (Datadog, Snowflake, Linear, orchestrated harness) integrated and operational.
I Am Often Wrong
4 hours ago
- The author's problem-solving approach involves six iterative steps: understand available information, gather missing information, define the problem, define a clear and simple approach, define a goal, and act with urgency.
- New information often requires revisiting and redefining steps 3–5, making the process iterative and adaptable.
- The most common failure modes are failing to clearly define the problem and failing to define a clear and simple approach, leading to complex plans and unclear success criteria.
- The author values being wrong because it helps refine the problem, find the right solution, and accelerate learning.
- Real-time feedback is encouraged to help individuals and teams learn more quickly and avoid these failure modes.
A Necessary History of the Oddest Letter: W
3 hours ago
- The letter W originated in the 6th century as a doubled U (UU) in Frankish Latin texts to represent the Germanic /w/ sound, which had no Latin equivalent after the fall of Rome.
- In Old English, the runic letter wynn (Ƿ) was also used for /w/, but it was gradually replaced by W after the Norman Conquest due to French and Latin influence.
- The sound /w/ was present in ancient Latin and Greek but later lost or altered; English preserved it due to its migration to Britain.
- W also arose from earlier G in Middle English (e.g., fugol → fowl) and was used to indicate vowel shifts, such as in town and cow after the Great Vowel Shift.
- The letter W has a rounding effect on following vowels, as seen in was vs. has, a modern development not present in Shakespeare's time.
- English spelling inconsistently uses OW and OU for the same shifted vowel (e.g., now vs. hour), with a half-hearted rule placing OW at word ends.
- The name 'double U' reflects its origin from two Us, while French calls it 'double vé' due to the later split of V and U.
Samsung is expected to more than double output of its HBM4 and HBM4E DRAM
3 hours ago
- Samsung plans to more than double HBM4 family output next year, including HBM4 and HBM4E, with a 2.5-fold increase in glass carrier demand.
- Outsourced glass carrier cleaning volume will rise to 50,000 sheets per month in 2024 from 20,000 in 2023, up from 10,000 last year.
- Glass carriers prevent wafer bending during thinning and drilling, crucial for high-stack HBM production.
- HBM4 and HBM4E center on 12-layer and higher stacks, with output likely growing at least twofold due to increased glass carrier usage.
- Samsung started mass production of HBM4 in February and provided 12-layer HBM4E samples to Nvidia in May.
- Industry expects Samsung's HBM production scale to grow nearly 40% to ~250,000 wafers (monthly input) next year.
- HBM4 family share is projected to rise from ~40% this year to ~80% next year as HBM4E mass production ramps up.
The Patriots' kicker scoreboard trick debunked
7 hours ago
- Patriots had an unusually large +9.4% field goal advantage over opponents from 2001-2019.
- Alleged trick: Gillette Stadium's video board showed reverse-angle shots, creating a second set of goalposts for opposing kickers.
- Statistical modeling shows opponents performed worse both at Foxboro and on the road against the Patriots, indicating other factors like special teams and game pressure.
- The specific effect attributable to the Foxboro board is only about 1.4%, with a standard error of 3.1%, making it statistically insignificant.
- The stadium's worst ranking among NFL venues is within the range expected by random chance in simulations.
- Installation of larger HD boards in 2010 did not increase the supposed effect; shortfall actually decreased.
- Conclusion: The scoreboard trick likely had negligible impact; professional kickers are adept at ignoring distractions.
Woman falsely charged with bank theft sues for $10M over AI facial recognition
4 hours ago
- Tennessee grandmother Angela Lipps was falsely accused of bank theft in North Dakota after AI facial recognition software matched her social media photo to a fake ID used by a suspect.
- She was arrested while babysitting and detained for six months, losing her residence, car, and other personal property before charges were dropped on Christmas Eve.
- Lipps’ attorney provided bank records proving she was in Tennessee during the crimes, and the suspect in surveillance footage did not resemble her in build, facial features, or tattoos.
- The $10 million federal lawsuit names the city of Fargo and detective Lucas Heck, alleging improper reliance on AI facial recognition, lack of thorough investigation, and inadequate police training.
- Fargo Police Chief acknowledged errors, stating the department assumed wrongly that surveillance photos were included, and has since prohibited use of the AI facial recognition system.
Overdue library magazine returned after 132 years
4 hours ago
- A man returned an 1894 magazine to Concord Public Library, 132 years overdue, after the library announced a fine freeze.
- Archivist Jennifer Needham handled the return, calling it a pleasant surprise and a piece of library history.
- The magazine is in good condition except for a cut-out illustration of Abraham Lincoln, and will be put on display later.
Autonomous strike drone uses Nvidia Jetson Orin Nano to pick and bomb targets
4 hours ago
- Autonomous NATO strike drone uses Nvidia Jetson Orin Nano to independently select and attack targets, requiring no human input or external communications.
- Swedish startup Scaleout Systems demonstrated the drone under BAE Systems' ALMA program, where a loitering munition autonomously detected, ranked, and struck an armored vehicle.
- All AI processing is done onboard, with manual input optional, and the pilot acting only as a failsafe; the mission completed in under 320 seconds without communication.
- Scaleout uses 'federated learning' to train AI models locally and share updates, part of NATO's DIANA accelerator and FEDAIR project, receiving €100,000 in funding.
- An arctic test at -18°C used an Airolit S1 with YOLOv8 Nano on Nvidia's Jetson Orin Nano, achieving 30 fps and 30 ms latency, with ranging via pinhole camera model.
- A separate Swedish Air Force test simulated communication loss, with the drone maintaining inference and active learning while offline, then backfilling data on reconnect.
- These demonstrations occur before any official NATO procurement order or combat deployment; a UN expert group recommends human control to comply with laws of war.
- The technology shows autonomous target selection on accessible hardware, but whether BAE moves ALMA to procurement remains undecided.

firstprev0next

Hasty Briefsbeta

The LLMs Yearn for the Spines

Why MCP Was Always a Bad Idea

What's been going on in w64devkit the past year

Frontier Labs Are Selling Garbage to Fools in Washington

Software Sandboxing: The Basics

Self-hosted inference orchestrators compared: LocalAI, exo, GPUStack, vLLM

The shape that does not return

Show HN: Three genlocked RP2350B make a console – 3k sprite pixels per line)

As White House shields the AI gold rush, Trump family and other allies strike...

US Revokes Limits on Power Plants' Climate Pollution

Trying the Software Factory Pattern

I Am Often Wrong

A Necessary History of the Oddest Letter: W

Samsung is expected to more than double output of its HBM4 and HBM4E DRAM

The Patriots' kicker scoreboard trick debunked

Woman falsely charged with bank theft sues for $10M over AI facial recognition

Overdue library magazine returned after 132 years

Autonomous strike drone uses Nvidia Jetson Orin Nano to pick and bomb targets