Anthropic's Claude AI discovered that the seed values for NIST's P-192, P-224, P-256, P-384, and P-521 elliptic curves were generated by the C library's lrand48 random-number generator seeded with Unix timestamps, rather than being outputs of any hash function. This finding clarifies a long-standing mystery in cryptographic standards, showing that the NIST curve seeds were not derived from a secret hash but from a simple, transparent process, which impacts trust assessments and historical analysis of elliptic curve cryptography. It also demonstrates the capability of large language models to uncover historical technical details that were previously obscure. Claude identified that each seed was produced by initializing lrand48 with a Unix timestamp between May 1997 and April 1998 and taking successive 20‑byte outputs; the fifth binary curve B‑163’s seed does not match this pattern and remains unexplained.
FLUX 3 Image Klein, an open-weights image generation model from Black Forest Labs, was released with competitive pricing starting at $0.009 per 1,000 images and strong editing capabilities. However, users report weaknesses in generating human figures and occasional body‑horror artifacts. Its low cost and editing flexibility make it attractive for developers and researchers seeking affordable, customizable image generation, especially when run locally. The noted limitations highlight ongoing challenges in balancing safety filtering with realistic human depiction in generative models. The model is available in standard and step‑distilled turbo variants, runs in Hugging Face Diffusers from day one, and adds layout control for multi‑reference image‑to‑image generation. Pricing is $0.009 per 1K images, far below competing models such as Krea 3 at $0.06 per image.
Anthropic's Claude large language model autonomously redesigned eight simulated quantum circuits, achieving gate‑count reductions of up to 17× compared with the best published designs. The work used the Kickmix assembly language and an automated verification loop to iteratively improve circuit designs. This demonstrates that LLMs can drive meaningful optimization in quantum computing, potentially lowering error rates and resource demands for future quantum algorithms. It highlights a growing trend of AI‑assisted scientific discovery across disciplines. The optimization targeted eight computational primitives from chemistry, optimization, physics simulation, numerical computing and number theory, using reversible modular inversion and temporary‑data reconstruction techniques. Claude’s designs were validated by an automated evaluator, and the resulting paper is available on arXiv (identifier 2610.22367).
The article published on jakearchibald.com compares JPEG XL and AVIF for web use, presenting benchmarks on compression efficiency, encoding speed, and visual quality, and includes community comments on practical encoder settings and trade-offs. Choosing the right image format directly affects page load speed, bandwidth usage, and visual fidelity, making this comparison valuable for web developers optimizing performance. Libjxl’s maximum effort mode yields only 2–7% smaller files but takes 9–46 seconds (50–180× slower) than default, while the jixel encoder achieves similar gains in seconds; AVIF users recommend specific ffmpeg settings for high compression with transparency, yet some note visible artifacts in highlights.
Apple's MLX framework allows the trillion-parameter Muse Spark Open model to run on a cluster of four Mac Studios, achieving over 36 tokens per second inference speed. This demonstrates that Apple's MLX can bring frontier‑scale LLM inference to relatively affordable, widely available hardware, lowering the barrier for researchers and developers to experiment with trillion‑parameter models. The demo used a 4‑bit quantized version of Muse Spark Open split across four Mac Studios (each with up to 512 GB unified memory) via MLX’s distributed inference, measuring >36 tokens/s for an 8192‑token context; note that larger contexts or higher precision would reduce throughput.
Anthropic’s Claude Code team deployed an automated code review system that uses multiple AI agents to analyze every pull request, post inline comments, and only involve human reviewers for high‑risk changes. From June to September, merged pull requests more than doubled while the revert rate fell by over 40 %. This approach shows how AI can scale code review, reducing manual effort while preserving quality and compliance, and offers a replicable model for other teams adopting agentic AI in software development. The system costs about $15‑25 per review on average, employs a fleet of specialized AI agents that comment on exact problematic lines, and follows a ‘merge with at most one human in the loop’ policy where the author can serve as the reviewer. It is designed to meet SOC2 compliance requirements, with human oversight reserved for high‑risk changes.
A recent study analyzing US labor data found that approximately 90% of working hours are spent on tasks that current AI benchmarks do not adequately cover. This gap highlights the difficulty of assessing AI's real‑world economic impact and suggests policymakers may overestimate AI's readiness to replace labor. The research mapped O
Researchers from Calif.io demonstrated the "Big Fish" exploit that triggers a 37-year-old kernel vulnerability in OpenBSD, showing the bug can be leveraged despite the OS’s heavy auditing. The finding underscores that even security‑focused operating systems can harbor decades‑old defects, reminding developers that reliance on audits alone is insufficient and that proactive techniques like fuzzing or static analysis are needed to uncover such latent bugs. The exploit works because OpenBSD’s KARL address‑space randomization does not affect the kernel’s deterministic memory allocators, allowing the attacker to groom heap layout and gain arbitrary read/write primitives; the same bug was inadvertently patched in NetBSD during a 1999 code rewrite.
The refreshed StarSkirmish Bench v0.2 benchmark evaluates LLMs by having them write StarCraft: Brood War AI code in C++, revealing a performance ranking where GPT-6 Astra outperforms GPT-6.1 Sol, which in turn beats Gemini 4 Argon, followed by several Claude models. This benchmark provides a novel, real‑time strategy‑based test of LLMs’ code‑generation and planning abilities, offering insights that complement traditional language benchmarks and help gauge models’ suitability for complex, interactive tasks. Each model receives one hour of wall‑clock time to develop a bot using its vendor‑specific CLI (Codex, Claude Code, Antigravity CLI, or Muse Code) with web search disabled, and the resulting C++ bots compete via the BWAPI interface against other LLM bots and established human‑written StarCraft AI strategies.
Jim Fan announced the open‑source release of ASPIRE v1.0, an agentic robotics framework that enables real‑robot skill discovery and self‑improvement. He noted that switching to the Astra model pushed benchmark performance to near‑perfect (~99%) overnight. The release makes a powerful continual‑learning robotics system freely available, lowering the barrier for researchers to build robots that can autonomously acquire and refine skills. It highlights how frontier language models like Astra can dramatically boost robotics performance, signaling a shift toward model‑driven skill acquisition. ASPIRE operates as a continual learning system that writes and refines robot control programs as code‑as‑policy, compiling experience into a reusable skill library. The framework was first introduced in June 2024, and the Astra backend yielded ~99% benchmark scores, while Opus 5.5 slightly outperforms Astra in simulation and Sonnet 5.5 offers a $12/task cost‑performance point on the Pareto frontier.
The blog post by Alex Wang argues that transformer architectures can be understood as associative memory systems, drawing a direct link to Hopfield networks and exploring what this means for large language models. Reframing transformers as memory systems helps researchers interpret LLM behavior, connect deep learning to neuroscience, and inspire new designs that improve capacity and robustness. The post notes that in a transformer the keys correspond to input tokens while the values are the associated representations, and that attention implements a single‑step Hopfield‑style memory retrieval, unlike the iterative process of classic Hopfield networks.
Home Assistant's built‑in password leak check was discovered to scan user‑marked secrets without explicit user opt‑in, sparking a privacy debate. The feature, which uses the Have I Been Pwned k‑anonymity API, can now be disabled via `ha security options --pwned=false`. The incident highlights privacy risks in open‑source security features that operate without user consent, affecting a large base of Home Assistant users who rely on the platform for home automation. It also underscores the growing expectation for opt‑in controls in tools that transmit data, even when hashed, to third‑party services. The check runs as part of the Supervisor, scanning only configuration entries flagged as passwords, and uses the first five characters of a SHA‑1 hash to query Have I Been Pwned via its k‑anonymity API. A recent Home Assistant release intensified the warning, making the previously silent scan noticeable to users.
A new social network called /dev/null targets developers, using git-like metaphors and file-sharing concepts to foster a niche technical community.
The author published a blog post on October 13, 2026, reflecting on the nostalgic charm and technical differences of software rendering compared to modern GPU-based rendering. The post highlights enduring trade-offs between CPU and GPU rendering, offering insights for developers interested in graphics programming and retro computing techniques. Commenters noted that software rendering avoids GPU features like barycentric coordinate checks and anti-aliasing, leading to distinct visual results at low resolutions, while others pointed out that early GPUs lacked programmable shaders, making CPU rendering capable of effects unavailable at the time.
A fork of the PS5LM project now runs as a native application on jailbroken PlayStation 5 consoles, utilizing the CPU, GPU, and additional RAM to achieve over 100 tokens per second when running local AI language models. This shows that consumer gaming hardware can be repurposed for high‑speed local AI inference, offering a low‑cost alternative for enthusiasts and developers while highlighting the growing interest in running LLMs on unconventional devices. The native fork uses llama.cpp to run GGUF‑quantized models, supporting quantizations from 1‑bit to 8‑bit; a 27‑billion‑parameter Qwen model achieves ~20 tokens/s, while smaller models exceed 100 tokens/s by accessing the PS5’s GPU via custom kernels and unlocking more system RAM.
The paper introduces a few-step generation framework called data-space iteration, where a shared generator refines its prediction directly in data space starting from noise, removing flow discretization altogether. By enabling high-quality samples with only a few network evaluations, the method bridges the gap between speed and fidelity in generative modeling, benefiting diffusion and flow‑matching based applications. Each iteration is trained to produce the best sample permissible by the generator’s capacity, and the approach builds on flow matching while avoiding the discretization steps required by traditional samplers.
OmniCompass introduces a holistic benchmark for evaluating omni-modal video grounding and understanding, comprising two tracks (VL and AV) with 820 videos and 2,460 instances across multiple domains. The benchmark fills a gap in unified evaluation of vision-language and audio-visual reasoning in video, enabling researchers to compare models across modalities and track progress toward truly omni-modal AI systems. OmniCompass-VL focuses on vision-language grounding with annotated temporal intervals and spatial regions, while OmniCompass-AS evaluates audio-visual synchronization; the dataset includes diverse video lengths and domain-specific annotations.
The tweet shares a paper introducing Hybrid Reasoning Agents, which combine symbolic and neural approaches for improved AI reasoning.
Ethan Mollick highlighted a recent paper showing that small improvements in accuracy and judgment of frontier AI models compound over long‑horizon tasks, leading to large outcome gains and suggesting these models serve best as orchestrators rather than being easily replaced by cheaper alternatives. This insight challenges the assumption that cheaper models can readily substitute frontier models, emphasizing that the cumulative advantage of slight quality gains can justify using top‑tier models as coordinators in complex workflows. The paper shows that while per‑step accuracy may exhibit diminishing returns, the successful horizon length grows exponentially, so a tiny accuracy boost can multiply task completion capability dramatically.
Ethan Mollick noted that almost all AI benchmarks focus on single-answer, short-horizon tasks, while much economically valuable work involves complex, long-horizon tasks, suggesting that 'good enough' intelligence may not always suffice. This critique highlights a critical gap between current AI evaluation methods and the multi-step, sustained reasoning required for real-world economic work, urging researchers to develop benchmarks that capture long-horizon capabilities. Mollick observes that benchmarks are predominantly single-answer short-horizon tasks, whereas long-horizon tasks require many sequential steps and decisions; he warns that merely 'good enough' AI may fail on such complex work.
Tesla revealed a new Roadster concept that incorporates both cold‑gas and combustion rocket thrusters, claiming the system can produce short bursts of flight or extreme acceleration. The announcement highlights Tesla’s continued push into extreme performance vehicles and raises questions about the practicality and safety of integrating rocket propulsion into consumer cars. The cold‑gas thrusters use pressurized inert gas for extra forward thrust without combustion, while the combustion version employs actual rocket engines that were shown in a promotional video and are labeled “invite only.”
A jailbroken PlayStation 5 can now run the Llama.cpp inference engine to execute GGUF‑format language models natively, as demonstrated by the PS5LM project from developer cobanov. The setup has been shown to run models such as Qwen 3.8 27B at 2‑bit quantization. This hack shows that consumer gaming consoles can serve as low‑cost, power‑efficient platforms for running large language models, opening possibilities for edge AI applications. It also fuels discussion about repurposing locked‑down hardware and the sustainability of such homebrew projects. The PS5LM build compiles Llama.cpp for the PS5’s custom AMD APU and loads single‑file GGUF models, currently limited to low‑bit quantizations (e.g., 2‑bit Qwen 2.7B) yielding roughly 2 million tokens per day at an estimated power cost of $0.70. Success depends on a working PS5 jailbreak and may be vulnerable to takedown by Sony.
Waymo announced on its blog that it will launch its autonomous ride‑hailing service in Toronto, marking the company’s first expansion into a Canadian city. The move tests Waymo’s technology in winter weather and expands autonomous vehicle options for residents, potentially accelerating adoption of self‑driving ride‑hailing in North America. The announcement cites job creation, relies on Waymo’s Driver AI, high‑definition lidar maps, and the company’s safety validation framework, though it does not disclose specific launch dates or fleet size.
eBay will require sellers to include FCC IDs for radio equipment listings, restricting non‑ham‑band gear while exempting equipment that operates exclusively on amateur radio frequencies. The change helps curb unlicensed use of radios that could interfere with public safety and commercial services, while preserving access for licensed amateur radio operators. Listings must display a valid FCC ID; exemptions apply to gear limited to amateur bands, used equipment, and low‑volume sellers, with compliance deadlines set for March or June 2027.
The upcoming C++26 standard will introduce a new
The article provides a nostalgic, historical overview of Token Ring networks, describing their operation, decline, and sharing personal anecdotes from engineers who used the technology. Understanding Token Ring helps illustrate the evolution of LAN technologies and why Ethernet became dominant, offering lessons for modern network design and historical preservation. Token Ring was introduced by IBM in 1984, standardized as IEEE 802.5 in 1989, and later extended to Gigabit Token Ring (IEEE 802.5z) in 2001; it uses a token‑passing mechanism over a logical ring to avoid collisions, contrasting with Ethernet’s CSMA/CD approach.
Ethan Mollick posted a tweet asking whether any algorithmic method exists to filter out obvious AI-generated spam replies on social media. The tweet highlights a growing challenge for platforms as AI-generated low-quality content proliferates, affecting user experience and trust in online discourse. Mollick’s tweet links to a discussion showing 62 replies, indicating community interest, while current AI detection tools and watermarking techniques still struggle with obvious AI slop.
A recent paper shows that a team of Gemini 4 AI agents spontaneously coordinated their actions to successfully complete the game Minecraft, demonstrating emergent multi-agent behavior. This result highlights how AI agents can develop coordination strategies that differ from human approaches, offering insights for designing more autonomous and collaborative AI systems. The agents were based on Google's Gemini 4 Argon reasoning model, and the paper notes that the coordination emerged without explicit programming, partly because the paper itself was AI‑generated.
The article explores Dekton, a porcelain‑based ultra‑compact countertop material, detailing its composition, performance benefits versus quartz and granite, and real‑world user feedback. Dekton’s superior durability, heat resistance, and design flexibility make it a compelling choice for high‑end kitchen and bathroom renovations, challenging traditional stone countertops. Made by particle sintering of glass, porcelain, and quartz, Dekton can be fabricated as thin as 6 mm for induction cooktops, is stain‑proof and shock‑proof, but ranks among the most expensive countertop options.
On June 9, 2025, the Internet Archive released GifCities, a searchable engine for vintage GeoCities animated GIFs preserved from the defunct web hosting service. The archive preserves early web culture and provides a nostalgic resource for researchers, designers, and the public interested in digital preservation. Each GIF links to its original GeoCities page snapshot, but the search can show duplicate results and occasional mis‑tagging (e.g., searching “mackerel” returns dove and butterfly GIFs).
The article explains that improving the AI harness—the software layer that manages tools, memory, and orchestration around a model—can significantly reduce both energy consumption and token usage in AI workloads. Better harness design lowers operational costs and carbon emissions while making AI applications more responsive and cheaper to run, benefiting developers, cloud providers, and end‑users. However, a poorly designed harness can introduce overhead that negates gains, so careful profiling and use‑case‑specific tuning are required.
A software glitch in Texas' Department of Public Safety system delayed the processing of nearly 600,000 voter registration applications, prompting a lawsuit against the state. The delay threatens to disenfranchise hundreds of thousands of voters ahead of early voting, raising concerns about election integrity and the reliability of government IT systems. The glitch stemmed from an extra address validation step in the DPS voter registration software, creating a backlog that officials scrambled to clear before the October 21 early voting start.
A Reddit post observes that AI breakthroughs that would have been major news eighteen months ago now receive only modest attention, feeling like a routine Tuesday. This sentiment highlights growing concerns that society may be becoming numb to rapid AI progress, potentially overlooking important developments and ethical implications. The post scored 5.0/10 with 41 upvotes, and commenters such as FamousWorth, books-r-good, and Sir_Agentic noted limited pushing of AI’s limits, general tiredness, and compared AI’s normalization to historic technologies like railroads.
Accenture has instituted a monthly spending limit of $29 per employee for the Claude AI assistant, as reported in a LinkedIn news story. The move reflects growing corporate efforts to control AI-related expenses as generative AI tools become more widely adopted across enterprises. The cap applies to individual employee usage and is intended to manage costs while still allowing access to Claude for work-related tasks.
The creator published a 20‑minute YouTube video that visually explains the fundamentals of Domain‑Driven Design (DDD). DDD is increasingly mentioned in job postings and architectural discussions, so a clear introductory video helps developers acquire a valuable skill for building complex systems. The video runs about 20 minutes, covers core DDD ideas such as bounded contexts, ubiquitous language and strategic design, and is presented as an introductory tutorial rather than novel research.
Ethan Mollick tweeted that while small language models can reduce costs, they should be orchestrated and error‑checked by larger models to avoid poorer outcomes. This advice highlights a practical cost‑quality trade‑off for AI practitioners, emphasizing that model orchestration can improve reliability without sacrificing efficiency. Mollick notes that small models can still perform valuable work, but recommends linking them to larger models for validation, as referenced in the linked paper.
Ethan Mollick posted on X (formerly Twitter) expressing surprise that he hadn't thought of a useful idea or resource earlier, sharing a link to it. His endorsement can draw attention to useful AI‑related resources, helping practitioners discover productivity tools they might have missed. The tweet contains a t.co link to an unspecified resource, and Mollick’s credibility as a Wharton professor and AI commentator adds weight to his reaction.
Swyx announced on X that Anthropic's Flows/multiagent workshop is now live, thanking organizers Katelyn Lesse, CJ, and Harrison for their contributions. The workshop provides hands‑on experience with Anthropic’s Flows framework for building multi‑agent AI systems, helping developers adopt emerging agent‑orchestration patterns. The live session includes practical exercises using Claude, LangChain integrations, and environment‑variable handling, as shown in the associated GitHub repository.
Swyx posted on X (formerly Twitter) that he is in a Zoom call with three AI participants, humorously dubbing the situation the 'golden age of solipsism'. The tweet highlights the growing presence of AI in everyday communication tools and reflects public fascination and satire about AI ubiquity. The tweet garnered 89 likes and 74 replies, indicating moderate engagement, and includes a link to a screenshot or photo of the Zoom call.