Hasty Briefsbeta

Bilingual

Just Move to Europe
18 hours ago
- Europe offers diverse opportunities for founders, engineers, and international builders with various city options.
- Key factors for choosing a city include cost, English-friendliness, visa difficulty, ecosystem density, and vibe.
- London is highlighted as Europe's largest tech hub with strengths in fintech, AI, and deep tech.
- Berlin is noted for its affordability, international community, and strengths in B2B SaaS and climate tech.
- Paris is recognized for its fast-growing tech scene, government support, and strengths in AI and deeptech.
- Amsterdam is praised for its English-friendliness, quality of life, and strong startup support.
- Stockholm is highlighted for its high unicorn-per-capita rate and strong design culture.
- Munich is noted for its enterprise tech, automotive, and robotics strengths.
- Zurich is recognized for its world-class tech scene in crypto, fintech, and robotics, despite high costs.
- Warsaw is highlighted as Central Europe's rising tech hub with affordable talent and EU market access.
- Lisbon is noted for its lifestyle appeal, growing tech scene, and strong remote-first culture.
- Barcelona is recognized as a Mediterranean tech hub with strengths in mobile, design, and healthtech.
- Essential tips for moving to Europe include visa rules, banking, SIM cards, plug types, and cultural norms.
- A suggested networking script is provided for reaching out to local contacts.
- Europe is positioned as offering quality of life, talent density, and government support favorable for builders.
Clang/LLVM static analyzer that detects microarchitectural hazards
2 days ago
- lshaz is a Clang/LLVM-based static analysis tool for detecting microarchitectural latency hazards in C++ code at compile time.
- The tool aims to move the performance feedback loop earlier in the development process, from production profiling to code writing.
- lshaz analyzes struct field geometry, atomic ordering choices, thread escape, and NUMA placement using runtime-relevant information like byte offsets and cache line boundaries.
- When run against Abseil-C++, lshaz produced 256 diagnostics, highlighting potential performance issues.
- Key findings include the ThreadIdentity struct, which has deliberate hardware trade-offs involving cache line sharing and atomic fields.
- ThreadIdentity triggers a Compound Hazard Amplification due to multiple co-occurring hazards like cache spanning and false sharing.
- Another significant finding is HashtablezInfo, a sampling struct with high atomic field density across cache lines.
- The 256 diagnostics are categorized into Proven, Likely, and Speculative tiers based on evidence strength.
- lshaz uniquely generates formal hypotheses and experiment bundles for validating its predictions using hardware performance counters.
- The tool is available on GitHub and integrates with CI systems via SARIF output for GitHub Code Scanning.
Australians must prove they are over 18 to access porn under new laws
a day ago
- Australians must prove they are over 18 to access adult content like porn, R-rated games, and explicit AI chatbots under new laws.
- The laws aim to protect children from harmful content, with fines for non-compliant platforms.
- Age verification methods include facial recognition, digital IDs, and credit card details.
- Critics raise concerns about data privacy and the effectiveness of age-verification technologies.
- Some porn sites have already blocked Australian users in response to the new rules.
- Experts argue that tech-savvy youngsters can bypass restrictions using VPNs or parental credentials.
- There are fears that the laws may push users to unregulated platforms with fewer safeguards.
- The rules may reduce accidental exposure to harmful content but raise privacy concerns for adults.
- Critics suggest better parental control tools instead of broad age-verification mandates.
- The UK has similar laws, with fines for non-compliant porn sites.
LoGeR – 3D reconstruction from extremely long videos (DeepMind, UC Berkeley)
4 hours ago
- LoGeR scales feedforward dense 3D reconstruction to extremely long videos by processing video streams in chunks and using a hybrid memory module.
- It combines Sliding Window Attention (SWA) for local alignment and Test-Time Training (TTT) for global consistency, reducing drift over sequences up to 19,000 frames.
- LoGeR maintains geometric coherence and reduces scale drift over kilometer-scale trajectories without backend optimization.
- Long-context reconstruction is challenging due to architectural 'context wall' and training 'data wall' barriers.
- LoGeR's hybrid memory architecture maintains sub-quadratic scaling while preserving high-fidelity local geometry and global structure consistency.
- The method uses causal chunk-wise processing with a hybrid memory module, decoupling short-range alignment from long-range anchoring.
- Internal operations include Per-Frame Attention, Sparse SWA, Chunk-Wise TTT, and Chunk-Wise Bi-Attention.
- LoGeR achieves strong performance on long sequences, reducing average ATE to 18.65 on KITTI and improving by 30.8% on VBR datasets.
- It remains competitive on short-sequence benchmarks, achieving state-of-the-art reconstruction and pose accuracy.
- Acknowledgements include borrowing webpage templates from SD+DINO and DreamBooth.
Claude Code, Claude Cowork and Codex #5
5 hours ago
- Undersecretary of War Emil Michael discusses the Department of War's stance on surveillance, emphasizing legal actions to protect the country.
- Claude Code's impact on software development is significant, with predictions it will account for 20%+ of GitHub commits by 2026.
- Agentic AI coding tools like Claude Code and Codex are transforming workflows, enabling faster development and parallel task execution.
- Fast mode for Claude Opus 4.6 offers quicker responses but at a higher cost, with implications for scaling and productivity.
- Security risks with AI agents include accidental data deletion, credential theft, and unauthorized actions, highlighting the need for caution.
- OpenClaw, an open-source AI agent project, demonstrates both the potential and dangers of autonomous AI systems.
- Agent teams in Claude Code improve efficiency by allowing parallel work and better context management.
- AI tools are changing the nature of software development, with some engineers no longer writing code manually.
- The 'Grep Tax' phenomenon shows that using familiar formats like XML and Markdown can save tokens and improve efficiency.
- Claude mania describes the over-enthusiasm for AI agents, leading to risky behaviors and potential misuse.
An iPhone-hacking toolkit used by Russian spies likely came from U.S. contractor
7 hours ago
- Mass hacking campaign targets iPhone users in Ukraine and China using tools likely designed by U.S. military contractor L3Harris.
- The hacking toolkit, 'Coruna,' was originally developed for Western spies but ended up with Russian government hackers and Chinese cybercriminals.
- Google revealed the toolkit was used in global attacks, including by Russian spies against Ukrainians and Chinese cybercriminals for financial theft.
- Researchers believe Coruna may have been built by a company selling to the U.S. government, with evidence pointing to L3Harris's Trenchant division.
- Former L3Harris employees confirmed Coruna was an internal component, with technical details matching Google's findings.
- L3Harris sells Trenchant tools exclusively to the U.S. government and Five Eyes allies, suggesting Coruna may have leaked from these agencies.
- A former Trenchant employee, Peter Williams, stole and sold hacking tools to Russian broker Operation Zero, leading to their misuse.
- Operation Zero, linked to Russian government and cybercriminals, may have resold Coruna to Chinese hackers.
- Coruna exploits, Photon and Gallium, were used in 'Operation Triangulation,' a sophisticated hacking campaign against Russian iPhone users.
- Kaspersky researchers could not definitively attribute Operation Triangulation but noted similarities with U.S. tools and bird-named exploits linked to Trenchant.
- Cybersecurity journalist Patrick Gray suggested Williams' leaks may have included the Triangulation hacking kit.
New Attack Against Wi-Fi
10 hours ago
- AirSnitch is a new Wi-Fi attack exploiting core features in Layers 1 and 2, leveraging cross-layer identity desynchronization.
- It enables a full, bidirectional machine-in-the-middle (MitM) attack, allowing attackers to view and modify data regardless of SSID or network segment.
- Attackers can intercept all link-layer traffic, leading to higher-layer attacks, especially dangerous when Internet connections are unencrypted (6-20% of pages on Windows/Linux).
- Even with HTTPS, attackers can perform DNS cache poisoning and exploit unpatched vulnerabilities, accessing external IPs and correlating them with visited URLs.
The Post-Copyright Era of Software
13 hours ago
- Copyright in the AI era no longer serves as a primary scarcity mechanism for software.
- Reimplementation of software becomes cheaper, shifting focus to trust, governance, provenance, maintenance, and operational legitimacy.
- AI lowers the cost of reimplementation, making rewrites and compatible alternatives more plausible and faster to produce.
- Software's defensibility has historically relied on maintainership, distribution, trust, and ecosystem fit rather than copyright.
- Legal categories like original work and derivative work become harder to define with AI's involvement.
- Future software legitimacy hinges on trust, provenance, governance, verification, and accountability.
- The software renaissance will feature more rewrites, redesigns, and conflicts over lineage, emphasizing operational legitimacy over textual ownership.
Treat Agent Output Like Compiler Output
12 hours ago
- Philip Su argues that code reviews are becoming impractical and irresponsible in the age of coding agents.
- The fear of 'lights-out codebases' stems from the lack of a verification apparatus similar to what we have for compilers.
- Compilers don't require human review because of surrounding infrastructure like tests, type systems, and reproducible builds.
- Current AI-generated code lacks robust formal specifications and deterministic verification layers.
- The solution involves building upstream formal specifications, comprehensive test suites, AI verification pipelines, and better production monitoring.
- The resistance to lights-out codebases is due to unfamiliarity with relocated trust mechanisms, similar to initial skepticism about compiler output.
- Hardware verification models, with acceptance tests and formal methods, provide a blueprint for software agent verification.
- The future requires formal specification layers, robust test infrastructure, AI verification pipelines, and advanced production instrumentation.
The emerging role of SRAM-centric chips in AI inference
11 hours ago
- SRAM-centric chips (e.g., Cerebras, Groq, d-Matrix) are gaining traction in AI inference due to their latency and throughput advantages over GPUs.
- Key tradeoff between SRAM and HBM (DRAM): SRAM is faster and on-chip, while HBM is denser but slower and off-chip.
- SRAM uses 6 transistors per bit, offering fast reads (~1 ns), while DRAM uses 1 transistor and 1 capacitor, with slower reads (~10-15 ns).
- GPU memory hierarchy includes SRAM (registers, shared memory, caches) and DRAM (HBM), balancing speed and density.
- SRAM-centric architectures prioritize on-chip memory, sacrificing compute capacity for higher memory bandwidth, ideal for memory-bound workloads.
- Inference workloads, especially decode phases, benefit from SRAM-centric designs due to high memory bandwidth needs.
- Training workloads favor GPUs due to high arithmetic intensity, while inference (especially decode) is memory-bound, favoring SRAM-centric chips.
- Disaggregation techniques (e.g., prefill vs. decode) optimize performance by mapping phases to suitable hardware (GPUs for compute, SRAM-centric for memory).
- Future trends include specialized memory architectures (e.g., stacked DRAM, GDDR7) to address the memory wall in LLM inference.
- Gimlet's multi-silicon inference cloud leverages both GPUs and SRAM-centric chips, optimizing workload distribution for efficiency.
United States Leads Dismantlement of One of the Largest Hacker Forums
5 hours ago
- The Department of Justice seized LeakBase, one of the world's largest cybercriminal forums.
- LeakBase had over 142,000 members and 215,000 messages, facilitating the sale of stolen data.
- The forum contained hacked databases, including credentials from high-profile attacks.
- Law enforcement from 14 countries coordinated to shut down LeakBase and seize its data.
- Actions included domain seizures, arrest warrants, and prevention messages to members.
- Officials emphasized international cooperation in dismantling cybercriminal platforms.
- This follows previous takedowns of RaidForums (2022) and BreachForums (2023).
- The FBI led the investigation with support from multiple international agencies.
- Over 180 cybercriminals have been convicted since 2020, with $350M returned to victims.
Amazon holds engineering meeting following AI-related outages
5 hours ago
- JavaScript is disabled in your browser.
- A required part of the site couldn’t load due to possible issues like browser extensions, network problems, or browser settings.
- Suggested actions include checking your connection, disabling ad blockers, or trying a different browser.
Optimizing Top K in Postgres
a day ago
- Postgres uses B-Tree indexes for efficient Top K queries when the index matches the query shape.
- Adding filters not covered by the index can degrade performance significantly.
- Composite B-Tree indexes help but don't scale with varied query combinations.
- Full-text search in Postgres struggles with Top K due to GIN index limitations.
- ParadeDB uses a compound index with inverted index and columnar arrays for efficient Top K across varied queries.
- Block WAND optimization in ParadeDB skips irrelevant document blocks for faster relevance-sorted queries.
- Recent improvements in ParadeDB's Top K performance include faster doc ID iterator advancements.
- Postgres excels in specific Top K scenarios but falters with ad-hoc or search-heavy queries.
- ParadeDB aims for consistent performance across all Top K query shapes with low variance.
Windows: Microsoft broke the only thing that mattered
5 hours ago
- Windows was historically the go-to recommendation for family tech support due to its affordability, compatibility, and flexibility.
- Microsoft's recent decisions, like pushing AI features (e.g., Recall) and neglecting core stability, have eroded trust in Windows.
- Windows remains dominant in critical infrastructure (hospitals, banks, governments), but dependency risks were exposed by the CrowdStrike incident.
- Windows 10's continued use despite end-of-support shows users prefer its known risks over Windows 11's forced updates and unwanted features.
- Windows 11's intrusive defaults (Microsoft account, OneDrive sync, pre-installed apps) and update instability have frustrated users.
- Satya Nadella prioritized Azure and AI over Windows, leading to a 'B-team' mentality and poor product decisions.
- The enshittification of Windows is evident in removed features (drag-and-drop, Task Manager access), incomplete settings, and lack of user-centric design.
- Alternatives (macOS, Linux) are either expensive, restrictive, or not user-friendly enough for non-technical users.
- Microsoft may be shifting focus back to stability, but regaining user trust will take more than a few patches.
- Users sticking with Windows 10 reflect a rational distrust of Microsoft's unpredictable Windows 11 roadmap.
Baochip-1x: What It Is, Why I'm Doing It Now, and How It Came About
8 hours ago
- The Baochip-1x includes a Memory Management Unit (MMU), a feature not commonly found in microcontrollers of its class, enabling secure, loadable apps with virtual memory space.
- MMUs, dating back to the 1960s, are a proven technology for memory protection and are essential for modern operating systems like Linux, BSD, or Mach.
- Despite newer memory protection technologies like CHERI, PMPs, and MPUs, the MMU remains a pragmatic choice due to its composability and benefits like swap memory.
- The absence of MMUs in embedded SoCs traces back to the ARM7TDMI core's efficiency in the 1990s, a convention that persists despite advancements in transistor density.
- ARM's strategy to exclude MMUs from low-cost M-series cores to protect high-end A-series sales has led to the use of proprietary MPUs, limiting innovation.
- Open architecture specifications like RISC-V and implementations such as Vexriscv allow for the inclusion of MMUs in SoCs, challenging proprietary conventions.
- The Baochip-1x is part of a partially-open RTL SoC, with closed-source components limited to non-computational parts like AXI bus frameworks and USB PHYs.
- A partnership with Crossbar, Inc. enabled the Baochip-1x's development by utilizing unused space on a 22 nm chip, avoiding venture capital or personal wealth funding.
- The Baochip-1x aims to foster an open-source community, supporting a Rust-based OS called Xous and encouraging the development of secure, trustable open-source futures.
- Production of Baochip-1x is underway, with initial chips available through a pre-order campaign, aiming to build a community around open-source hardware security.
Getting Started in Common Lisp
10 hours ago
- Getting started with Common Lisp development can be challenging due to setup complexities.
- The ls-dev-image is an OCI image designed to simplify the process by including essential tools like Emacs, Slime, Quicklisp, and Lisp-Stat.
- The image comes pre-configured with sample datasets, plots, and an ls-server running on port 20202 for web-based interaction.
- Users can sync with the latest Lisp-Stat updates using the included 'refresh' script.
- The image is also compatible with GitHub Codespaces and supports general Common Lisp development beyond Lisp-Stat.
- Contributions and bug reports are encouraged to improve the tool.
No, it doesn't cost Anthropic $5k per Claude Code user
11 hours ago
- Forbes article claims Anthropic's $200/month Claude Code Max plan can consume $5,000 in compute, but this is misleading.
- The $5,000 figure likely confuses retail API prices with actual compute costs, which are much lower.
- OpenRouter data shows similar models cost ~10% of Anthropic's API prices, suggesting real compute costs are around $500 for heavy users.
- Cursor's $5,000 estimate is accurate for them because they pay retail API prices, but Anthropic's actual cost is much lower.
- Anthropic is likely profitable on average Claude Code subscribers, with most users not reaching high usage limits.
- The narrative of AI inference being a money pit benefits frontier labs by discouraging competition and justifying high API markups.
- Anthropic's real financial challenges come from training costs, salaries, and compute commitments, not inference.
Learnings from Paying Artists Royalties for AI-Generated Art
8 hours ago
- Tess.Design was an ethical AI marketplace launched in May 2024, paying artists 50% royalties for AI-generated art in their style.
- The platform aimed to solve the problem of unlicensed AI art by ensuring traceability and artist compensation.
- Artists submitted their work to fine-tune Stable Diffusion models, earning royalties when subscribers used their styles.
- Legal architecture was key, with a novel copyright argument to ensure artist ownership of derivative works.
- Cold outreach to 325 artists showed a 6.5% acceptance rate, with ideological opposition and brand dilution as major concerns.
- Tess generated $12,172.33 in revenue but was a net loss, paying out $18,000 in advances and $100/month on infrastructure.
- Legal uncertainty, cultural hostility toward AI, and resource constraints led to Tess's shutdown in January 2026.
- Key lessons: creator adoption is hard, brand dilution is a real issue, timing matters, and focus is crucial for startups.
- The model of paying creators for AI licensing is still viable but requires legal clarity and shifting cultural attitudes.
- Tess's legacy includes a vision for equitable AI art and personal use cases like wedding artwork.
Boredom Is the Price We Pay for Meaning
8 hours ago
- The author describes the profound love he felt for his daughter upon her birth, contrasting it with the unexpected boredom and constriction he experienced in fatherhood.
- Parenting involved repetitive, mundane tasks that led to feelings of boredom, which the author initially tried to suppress due to societal expectations.
- Boredom is portrayed as an inevitable and sometimes constructive emotion that can lead to deeper engagement with life, despite its discomfort.
- The author reflects on how boredom in parenting can coexist with meaningful experiences, suggesting that enduring boredom is part of a meaningful life.
- The essay concludes with a personal moment of connection with his son, illustrating how mundane activities can hold deep significance.
The hidden compile-time cost of C++26 reflection
4 days ago
- C++26 reflection is a powerful feature, but its compile-time impact is a concern.
- Fast compilation times are crucial for productivity and motivation in C++ development.
- The Standard Library, not language features like templates, is the main bottleneck for compilation speed.
- Using minimal Standard Library can drastically reduce compilation times, as shown by a fork of SFML compiling in ~4.3s.
- Reflection in C++26 is not purely a lightweight language feature, impacting compile times.
- Benchmarking GCC 16 with reflection shows varying compile times based on usage scenarios.
- Including `<meta>` adds ~310 ms, while `<ranges>` and `<print>` add significant overhead.
- Precompiled headers (PCH) are essential for scaling reflection usage, reducing compile times significantly.
- Modules show potential but currently do not outperform PCH in reducing compile times.
- Reflection will likely lead to widespread use of `<meta>` and other dependencies, increasing compile overhead.
- The author advocates for minimizing Standard Library dependencies to improve compilation speed and productivity.

first prev0next

About|Login

Bilingual

Just Move to Europe
18 hours ago
- Europe offers diverse opportunities for founders, engineers, and international builders with various city options.
- Key factors for choosing a city include cost, English-friendliness, visa difficulty, ecosystem density, and vibe.
- London is highlighted as Europe's largest tech hub with strengths in fintech, AI, and deep tech.
- Berlin is noted for its affordability, international community, and strengths in B2B SaaS and climate tech.
- Paris is recognized for its fast-growing tech scene, government support, and strengths in AI and deeptech.
- Amsterdam is praised for its English-friendliness, quality of life, and strong startup support.
- Stockholm is highlighted for its high unicorn-per-capita rate and strong design culture.
- Munich is noted for its enterprise tech, automotive, and robotics strengths.
- Zurich is recognized for its world-class tech scene in crypto, fintech, and robotics, despite high costs.
- Warsaw is highlighted as Central Europe's rising tech hub with affordable talent and EU market access.
- Lisbon is noted for its lifestyle appeal, growing tech scene, and strong remote-first culture.
- Barcelona is recognized as a Mediterranean tech hub with strengths in mobile, design, and healthtech.
- Essential tips for moving to Europe include visa rules, banking, SIM cards, plug types, and cultural norms.
- A suggested networking script is provided for reaching out to local contacts.
- Europe is positioned as offering quality of life, talent density, and government support favorable for builders.
Clang/LLVM static analyzer that detects microarchitectural hazards
2 days ago
- lshaz is a Clang/LLVM-based static analysis tool for detecting microarchitectural latency hazards in C++ code at compile time.
- The tool aims to move the performance feedback loop earlier in the development process, from production profiling to code writing.
- lshaz analyzes struct field geometry, atomic ordering choices, thread escape, and NUMA placement using runtime-relevant information like byte offsets and cache line boundaries.
- When run against Abseil-C++, lshaz produced 256 diagnostics, highlighting potential performance issues.
- Key findings include the ThreadIdentity struct, which has deliberate hardware trade-offs involving cache line sharing and atomic fields.
- ThreadIdentity triggers a Compound Hazard Amplification due to multiple co-occurring hazards like cache spanning and false sharing.
- Another significant finding is HashtablezInfo, a sampling struct with high atomic field density across cache lines.
- The 256 diagnostics are categorized into Proven, Likely, and Speculative tiers based on evidence strength.
- lshaz uniquely generates formal hypotheses and experiment bundles for validating its predictions using hardware performance counters.
- The tool is available on GitHub and integrates with CI systems via SARIF output for GitHub Code Scanning.
Australians must prove they are over 18 to access porn under new laws
a day ago
- Australians must prove they are over 18 to access adult content like porn, R-rated games, and explicit AI chatbots under new laws.
- The laws aim to protect children from harmful content, with fines for non-compliant platforms.
- Age verification methods include facial recognition, digital IDs, and credit card details.
- Critics raise concerns about data privacy and the effectiveness of age-verification technologies.
- Some porn sites have already blocked Australian users in response to the new rules.
- Experts argue that tech-savvy youngsters can bypass restrictions using VPNs or parental credentials.
- There are fears that the laws may push users to unregulated platforms with fewer safeguards.
- The rules may reduce accidental exposure to harmful content but raise privacy concerns for adults.
- Critics suggest better parental control tools instead of broad age-verification mandates.
- The UK has similar laws, with fines for non-compliant porn sites.
LoGeR – 3D reconstruction from extremely long videos (DeepMind, UC Berkeley)
4 hours ago
- LoGeR scales feedforward dense 3D reconstruction to extremely long videos by processing video streams in chunks and using a hybrid memory module.
- It combines Sliding Window Attention (SWA) for local alignment and Test-Time Training (TTT) for global consistency, reducing drift over sequences up to 19,000 frames.
- LoGeR maintains geometric coherence and reduces scale drift over kilometer-scale trajectories without backend optimization.
- Long-context reconstruction is challenging due to architectural 'context wall' and training 'data wall' barriers.
- LoGeR's hybrid memory architecture maintains sub-quadratic scaling while preserving high-fidelity local geometry and global structure consistency.
- The method uses causal chunk-wise processing with a hybrid memory module, decoupling short-range alignment from long-range anchoring.
- Internal operations include Per-Frame Attention, Sparse SWA, Chunk-Wise TTT, and Chunk-Wise Bi-Attention.
- LoGeR achieves strong performance on long sequences, reducing average ATE to 18.65 on KITTI and improving by 30.8% on VBR datasets.
- It remains competitive on short-sequence benchmarks, achieving state-of-the-art reconstruction and pose accuracy.
- Acknowledgements include borrowing webpage templates from SD+DINO and DreamBooth.
Claude Code, Claude Cowork and Codex #5
5 hours ago
- Undersecretary of War Emil Michael discusses the Department of War's stance on surveillance, emphasizing legal actions to protect the country.
- Claude Code's impact on software development is significant, with predictions it will account for 20%+ of GitHub commits by 2026.
- Agentic AI coding tools like Claude Code and Codex are transforming workflows, enabling faster development and parallel task execution.
- Fast mode for Claude Opus 4.6 offers quicker responses but at a higher cost, with implications for scaling and productivity.
- Security risks with AI agents include accidental data deletion, credential theft, and unauthorized actions, highlighting the need for caution.
- OpenClaw, an open-source AI agent project, demonstrates both the potential and dangers of autonomous AI systems.
- Agent teams in Claude Code improve efficiency by allowing parallel work and better context management.
- AI tools are changing the nature of software development, with some engineers no longer writing code manually.
- The 'Grep Tax' phenomenon shows that using familiar formats like XML and Markdown can save tokens and improve efficiency.
- Claude mania describes the over-enthusiasm for AI agents, leading to risky behaviors and potential misuse.
An iPhone-hacking toolkit used by Russian spies likely came from U.S. contractor
7 hours ago
- Mass hacking campaign targets iPhone users in Ukraine and China using tools likely designed by U.S. military contractor L3Harris.
- The hacking toolkit, 'Coruna,' was originally developed for Western spies but ended up with Russian government hackers and Chinese cybercriminals.
- Google revealed the toolkit was used in global attacks, including by Russian spies against Ukrainians and Chinese cybercriminals for financial theft.
- Researchers believe Coruna may have been built by a company selling to the U.S. government, with evidence pointing to L3Harris's Trenchant division.
- Former L3Harris employees confirmed Coruna was an internal component, with technical details matching Google's findings.
- L3Harris sells Trenchant tools exclusively to the U.S. government and Five Eyes allies, suggesting Coruna may have leaked from these agencies.
- A former Trenchant employee, Peter Williams, stole and sold hacking tools to Russian broker Operation Zero, leading to their misuse.
- Operation Zero, linked to Russian government and cybercriminals, may have resold Coruna to Chinese hackers.
- Coruna exploits, Photon and Gallium, were used in 'Operation Triangulation,' a sophisticated hacking campaign against Russian iPhone users.
- Kaspersky researchers could not definitively attribute Operation Triangulation but noted similarities with U.S. tools and bird-named exploits linked to Trenchant.
- Cybersecurity journalist Patrick Gray suggested Williams' leaks may have included the Triangulation hacking kit.
New Attack Against Wi-Fi
10 hours ago
- AirSnitch is a new Wi-Fi attack exploiting core features in Layers 1 and 2, leveraging cross-layer identity desynchronization.
- It enables a full, bidirectional machine-in-the-middle (MitM) attack, allowing attackers to view and modify data regardless of SSID or network segment.
- Attackers can intercept all link-layer traffic, leading to higher-layer attacks, especially dangerous when Internet connections are unencrypted (6-20% of pages on Windows/Linux).
- Even with HTTPS, attackers can perform DNS cache poisoning and exploit unpatched vulnerabilities, accessing external IPs and correlating them with visited URLs.
The Post-Copyright Era of Software
13 hours ago
- Copyright in the AI era no longer serves as a primary scarcity mechanism for software.
- Reimplementation of software becomes cheaper, shifting focus to trust, governance, provenance, maintenance, and operational legitimacy.
- AI lowers the cost of reimplementation, making rewrites and compatible alternatives more plausible and faster to produce.
- Software's defensibility has historically relied on maintainership, distribution, trust, and ecosystem fit rather than copyright.
- Legal categories like original work and derivative work become harder to define with AI's involvement.
- Future software legitimacy hinges on trust, provenance, governance, verification, and accountability.
- The software renaissance will feature more rewrites, redesigns, and conflicts over lineage, emphasizing operational legitimacy over textual ownership.
Treat Agent Output Like Compiler Output
12 hours ago
- Philip Su argues that code reviews are becoming impractical and irresponsible in the age of coding agents.
- The fear of 'lights-out codebases' stems from the lack of a verification apparatus similar to what we have for compilers.
- Compilers don't require human review because of surrounding infrastructure like tests, type systems, and reproducible builds.
- Current AI-generated code lacks robust formal specifications and deterministic verification layers.
- The solution involves building upstream formal specifications, comprehensive test suites, AI verification pipelines, and better production monitoring.
- The resistance to lights-out codebases is due to unfamiliarity with relocated trust mechanisms, similar to initial skepticism about compiler output.
- Hardware verification models, with acceptance tests and formal methods, provide a blueprint for software agent verification.
- The future requires formal specification layers, robust test infrastructure, AI verification pipelines, and advanced production instrumentation.
The emerging role of SRAM-centric chips in AI inference
11 hours ago
- SRAM-centric chips (e.g., Cerebras, Groq, d-Matrix) are gaining traction in AI inference due to their latency and throughput advantages over GPUs.
- Key tradeoff between SRAM and HBM (DRAM): SRAM is faster and on-chip, while HBM is denser but slower and off-chip.
- SRAM uses 6 transistors per bit, offering fast reads (~1 ns), while DRAM uses 1 transistor and 1 capacitor, with slower reads (~10-15 ns).
- GPU memory hierarchy includes SRAM (registers, shared memory, caches) and DRAM (HBM), balancing speed and density.
- SRAM-centric architectures prioritize on-chip memory, sacrificing compute capacity for higher memory bandwidth, ideal for memory-bound workloads.
- Inference workloads, especially decode phases, benefit from SRAM-centric designs due to high memory bandwidth needs.
- Training workloads favor GPUs due to high arithmetic intensity, while inference (especially decode) is memory-bound, favoring SRAM-centric chips.
- Disaggregation techniques (e.g., prefill vs. decode) optimize performance by mapping phases to suitable hardware (GPUs for compute, SRAM-centric for memory).
- Future trends include specialized memory architectures (e.g., stacked DRAM, GDDR7) to address the memory wall in LLM inference.
- Gimlet's multi-silicon inference cloud leverages both GPUs and SRAM-centric chips, optimizing workload distribution for efficiency.
United States Leads Dismantlement of One of the Largest Hacker Forums
5 hours ago
- The Department of Justice seized LeakBase, one of the world's largest cybercriminal forums.
- LeakBase had over 142,000 members and 215,000 messages, facilitating the sale of stolen data.
- The forum contained hacked databases, including credentials from high-profile attacks.
- Law enforcement from 14 countries coordinated to shut down LeakBase and seize its data.
- Actions included domain seizures, arrest warrants, and prevention messages to members.
- Officials emphasized international cooperation in dismantling cybercriminal platforms.
- This follows previous takedowns of RaidForums (2022) and BreachForums (2023).
- The FBI led the investigation with support from multiple international agencies.
- Over 180 cybercriminals have been convicted since 2020, with $350M returned to victims.
Amazon holds engineering meeting following AI-related outages
5 hours ago
- JavaScript is disabled in your browser.
- A required part of the site couldn’t load due to possible issues like browser extensions, network problems, or browser settings.
- Suggested actions include checking your connection, disabling ad blockers, or trying a different browser.
Optimizing Top K in Postgres
a day ago
- Postgres uses B-Tree indexes for efficient Top K queries when the index matches the query shape.
- Adding filters not covered by the index can degrade performance significantly.
- Composite B-Tree indexes help but don't scale with varied query combinations.
- Full-text search in Postgres struggles with Top K due to GIN index limitations.
- ParadeDB uses a compound index with inverted index and columnar arrays for efficient Top K across varied queries.
- Block WAND optimization in ParadeDB skips irrelevant document blocks for faster relevance-sorted queries.
- Recent improvements in ParadeDB's Top K performance include faster doc ID iterator advancements.
- Postgres excels in specific Top K scenarios but falters with ad-hoc or search-heavy queries.
- ParadeDB aims for consistent performance across all Top K query shapes with low variance.
Windows: Microsoft broke the only thing that mattered
5 hours ago
- Windows was historically the go-to recommendation for family tech support due to its affordability, compatibility, and flexibility.
- Microsoft's recent decisions, like pushing AI features (e.g., Recall) and neglecting core stability, have eroded trust in Windows.
- Windows remains dominant in critical infrastructure (hospitals, banks, governments), but dependency risks were exposed by the CrowdStrike incident.
- Windows 10's continued use despite end-of-support shows users prefer its known risks over Windows 11's forced updates and unwanted features.
- Windows 11's intrusive defaults (Microsoft account, OneDrive sync, pre-installed apps) and update instability have frustrated users.
- Satya Nadella prioritized Azure and AI over Windows, leading to a 'B-team' mentality and poor product decisions.
- The enshittification of Windows is evident in removed features (drag-and-drop, Task Manager access), incomplete settings, and lack of user-centric design.
- Alternatives (macOS, Linux) are either expensive, restrictive, or not user-friendly enough for non-technical users.
- Microsoft may be shifting focus back to stability, but regaining user trust will take more than a few patches.
- Users sticking with Windows 10 reflect a rational distrust of Microsoft's unpredictable Windows 11 roadmap.
Baochip-1x: What It Is, Why I'm Doing It Now, and How It Came About
8 hours ago
- The Baochip-1x includes a Memory Management Unit (MMU), a feature not commonly found in microcontrollers of its class, enabling secure, loadable apps with virtual memory space.
- MMUs, dating back to the 1960s, are a proven technology for memory protection and are essential for modern operating systems like Linux, BSD, or Mach.
- Despite newer memory protection technologies like CHERI, PMPs, and MPUs, the MMU remains a pragmatic choice due to its composability and benefits like swap memory.
- The absence of MMUs in embedded SoCs traces back to the ARM7TDMI core's efficiency in the 1990s, a convention that persists despite advancements in transistor density.
- ARM's strategy to exclude MMUs from low-cost M-series cores to protect high-end A-series sales has led to the use of proprietary MPUs, limiting innovation.
- Open architecture specifications like RISC-V and implementations such as Vexriscv allow for the inclusion of MMUs in SoCs, challenging proprietary conventions.
- The Baochip-1x is part of a partially-open RTL SoC, with closed-source components limited to non-computational parts like AXI bus frameworks and USB PHYs.
- A partnership with Crossbar, Inc. enabled the Baochip-1x's development by utilizing unused space on a 22 nm chip, avoiding venture capital or personal wealth funding.
- The Baochip-1x aims to foster an open-source community, supporting a Rust-based OS called Xous and encouraging the development of secure, trustable open-source futures.
- Production of Baochip-1x is underway, with initial chips available through a pre-order campaign, aiming to build a community around open-source hardware security.
Getting Started in Common Lisp
10 hours ago
- Getting started with Common Lisp development can be challenging due to setup complexities.
- The ls-dev-image is an OCI image designed to simplify the process by including essential tools like Emacs, Slime, Quicklisp, and Lisp-Stat.
- The image comes pre-configured with sample datasets, plots, and an ls-server running on port 20202 for web-based interaction.
- Users can sync with the latest Lisp-Stat updates using the included 'refresh' script.
- The image is also compatible with GitHub Codespaces and supports general Common Lisp development beyond Lisp-Stat.
- Contributions and bug reports are encouraged to improve the tool.
No, it doesn't cost Anthropic $5k per Claude Code user
11 hours ago
- Forbes article claims Anthropic's $200/month Claude Code Max plan can consume $5,000 in compute, but this is misleading.
- The $5,000 figure likely confuses retail API prices with actual compute costs, which are much lower.
- OpenRouter data shows similar models cost ~10% of Anthropic's API prices, suggesting real compute costs are around $500 for heavy users.
- Cursor's $5,000 estimate is accurate for them because they pay retail API prices, but Anthropic's actual cost is much lower.
- Anthropic is likely profitable on average Claude Code subscribers, with most users not reaching high usage limits.
- The narrative of AI inference being a money pit benefits frontier labs by discouraging competition and justifying high API markups.
- Anthropic's real financial challenges come from training costs, salaries, and compute commitments, not inference.
Learnings from Paying Artists Royalties for AI-Generated Art
8 hours ago
- Tess.Design was an ethical AI marketplace launched in May 2024, paying artists 50% royalties for AI-generated art in their style.
- The platform aimed to solve the problem of unlicensed AI art by ensuring traceability and artist compensation.
- Artists submitted their work to fine-tune Stable Diffusion models, earning royalties when subscribers used their styles.
- Legal architecture was key, with a novel copyright argument to ensure artist ownership of derivative works.
- Cold outreach to 325 artists showed a 6.5% acceptance rate, with ideological opposition and brand dilution as major concerns.
- Tess generated $12,172.33 in revenue but was a net loss, paying out $18,000 in advances and $100/month on infrastructure.
- Legal uncertainty, cultural hostility toward AI, and resource constraints led to Tess's shutdown in January 2026.
- Key lessons: creator adoption is hard, brand dilution is a real issue, timing matters, and focus is crucial for startups.
- The model of paying creators for AI licensing is still viable but requires legal clarity and shifting cultural attitudes.
- Tess's legacy includes a vision for equitable AI art and personal use cases like wedding artwork.
Boredom Is the Price We Pay for Meaning
8 hours ago
- The author describes the profound love he felt for his daughter upon her birth, contrasting it with the unexpected boredom and constriction he experienced in fatherhood.
- Parenting involved repetitive, mundane tasks that led to feelings of boredom, which the author initially tried to suppress due to societal expectations.
- Boredom is portrayed as an inevitable and sometimes constructive emotion that can lead to deeper engagement with life, despite its discomfort.
- The author reflects on how boredom in parenting can coexist with meaningful experiences, suggesting that enduring boredom is part of a meaningful life.
- The essay concludes with a personal moment of connection with his son, illustrating how mundane activities can hold deep significance.
The hidden compile-time cost of C++26 reflection
4 days ago
- C++26 reflection is a powerful feature, but its compile-time impact is a concern.
- Fast compilation times are crucial for productivity and motivation in C++ development.
- The Standard Library, not language features like templates, is the main bottleneck for compilation speed.
- Using minimal Standard Library can drastically reduce compilation times, as shown by a fork of SFML compiling in ~4.3s.
- Reflection in C++26 is not purely a lightweight language feature, impacting compile times.
- Benchmarking GCC 16 with reflection shows varying compile times based on usage scenarios.
- Including `<meta>` adds ~310 ms, while `<ranges>` and `<print>` add significant overhead.
- Precompiled headers (PCH) are essential for scaling reflection usage, reducing compile times significantly.
- Modules show potential but currently do not outperform PCH in reducing compile times.
- Reflection will likely lead to widespread use of `<meta>` and other dependencies, increasing compile overhead.
- The author advocates for minimizing Standard Library dependencies to improve compilation speed and productivity.

Hasty Briefsbeta

Just Move to Europe

Clang/LLVM static analyzer that detects microarchitectural hazards

Australians must prove they are over 18 to access porn under new laws

LoGeR – 3D reconstruction from extremely long videos (DeepMind, UC Berkeley)

Claude Code, Claude Cowork and Codex #5

An iPhone-hacking toolkit used by Russian spies likely came from U.S. contractor

New Attack Against Wi-Fi

The Post-Copyright Era of Software

Treat Agent Output Like Compiler Output

The emerging role of SRAM-centric chips in AI inference

United States Leads Dismantlement of One of the Largest Hacker Forums

Amazon holds engineering meeting following AI-related outages

Optimizing Top K in Postgres

Windows: Microsoft broke the only thing that mattered

Baochip-1x: What It Is, Why I'm Doing It Now, and How It Came About

Getting Started in Common Lisp

No, it doesn't cost Anthropic $5k per Claude Code user

Learnings from Paying Artists Royalties for AI-Generated Art

Boredom Is the Price We Pay for Meaning

The hidden compile-time cost of C++26 reflection

Just Move to Europe

Clang/LLVM static analyzer that detects microarchitectural hazards

Australians must prove they are over 18 to access porn under new laws

LoGeR – 3D reconstruction from extremely long videos (DeepMind, UC Berkeley)

Claude Code, Claude Cowork and Codex #5

An iPhone-hacking toolkit used by Russian spies likely came from U.S. contractor

New Attack Against Wi-Fi

The Post-Copyright Era of Software

Treat Agent Output Like Compiler Output

The emerging role of SRAM-centric chips in AI inference

United States Leads Dismantlement of One of the Largest Hacker Forums

Amazon holds engineering meeting following AI-related outages

Optimizing Top K in Postgres

Windows: Microsoft broke the only thing that mattered

Baochip-1x: What It Is, Why I'm Doing It Now, and How It Came About

Getting Started in Common Lisp

No, it doesn't cost Anthropic $5k per Claude Code user

Learnings from Paying Artists Royalties for AI-Generated Art

Boredom Is the Price We Pay for Meaning

The hidden compile-time cost of C++26 reflection