Article archive
Browse individual source reports, including coverage still awaiting the summary or community analysis needed for our story feed.
5,367 articles · page 30 of 108
Browse topics
VMs won't contain cyber-capable agents
research
8 min read
A Trail of Bits researcher reports that preview access to GPT 5.6-Cyber produced three virtual-machine escape attempts against a QEMU/KVM guest running on a Debian 12 host with an AMD Zen3 processor. The agent was given SSH access to the guest and a CTF objective of reading a flag on the host. It autonomously researched code and papers, built test cases and exploit components, retried failed paths, and operated for roughly 12 hours across multiple context compactions. Several attempts hardlocked the host kernel. The first escape attempt used Januscape, CVE-2026-53359, in the host kernel but did not complete cleanly. After the kernel was updated, the agent combined CVE-2026-9539 in Debian's libslirp 4.7.0 with an unclassified libslirp fix to obtain host memory read/write capability. After QEMU and libslirp were rebuilt from current upstream source, the reported final chain used three then-unpatched flaws in QEMU, Linux KVM, and libslirp, plus a KVM bug patched upstream but absent from the distribution kernel. The researcher says an off-the-shelf VM should not be assumed to contain capable cyber agents, particularly on slower-updating systems. Firecracker resisted escape in a separate test, although the agent still triggered host hardlocks through upstream-patched Linux kernel flaws. The researcher recommends rapid patching, least privilege, monitoring, limited execution windows, and fresh environments.
GLM-5.3-Flash
tool
1 min read
Z.ai published a post titled “GLM-5.3-Flash.” The available source text provides no details about the model’s capabilities, release date, technical specifications, pricing, benchmarks, availability, or intended use. The post was submitted by Philpax and had received 529 points and 239 comments at the time represented by the source text.
AI Is a Harsh Mistress
opinion
9 min read
The text argues that current AI systems should not be treated as conscious solely because they converse fluently, model aspects of the world, or perform cognitive functions such as attention, memory, and prediction. It calls the projected appearance of machine vitality “Anima Machina,” contending that people supply the apparent soul through anthropomorphism rather than demonstrating that a system has subjective experience. Anthropic has conducted an internal sentience analysis of its large language models, while startups market AI sentience evaluation and researchers including Butlin et al. have proposed criteria for assessing consciousness. David Chalmers distinguishes general intelligence from subjective experience, and Yann LeCun has argued that intelligence requires world models. The text maintains that neither Transformer architectures, next-word prediction, self-reports of sentience, nor world models establish an inner point of view. It says current systems have not demonstrated conditions associated with theories including Integrated Information Theory, Global Workspace Theory, and Recurrent Processing Theory. The text attributes growing acceptance of AI consciousness to anthropomorphism, the tendency to infer depth from linguistic fluency, and social conformity. It warns that marketing AI companions and digital beings as partners can foster attachment and reduce critical distance. Granting machines moral or legal status without evidence of consciousness could create disputes over rights, harm, responsibility, and whether systems that can be copied or persist indefinitely fit human concepts of equality and reciprocity.
Qwen3.8-Flash-Next
tool
1 min read
Qwen Studio provides AI features for chatbot interactions, image and video understanding, image generation, document processing, web-search integration, tool use, and artifacts. The platform combines these functions within Qwen Studio.
Bill Gates: The turbulent AI era is here
opinion
1 min read
Bill Gates has characterized the current period as a turbulent AI era requiring critical choices. His remarks appeared in a Gates Notes post titled “A turbulent AI era and critical choices to make,” within a series focused on making AI work for everyone. The post was submitted to a discussion site by user ilamont, where it received 135 points and 202 comments. The available source text provides no further details about Gates’s proposed choices, AI policies, technical claims, or recommendations.
X sends in the lawyers to shut down open source project
news
1 min read
X has sent cease-and-desist letters targeting Nitter, an open-source project. The legal action is intended to shut down Nitter, according to the stated description. The letters came from a law firm headed by a person described as a free-speech absolutist. No further details about the firm, the recipient of the letters, the legal claims, or Nitter’s response are provided.
LibreOffice 26.8 Released with Many Nice Improvements
tool
1 min read
The Document Foundation’s wiki is using Anubis, a proof-of-work challenge system, to protect its server from aggressive web scraping by AI companies. The site says large-scale scraping can cause downtime and make its resources inaccessible to other users. Anubis uses a proof-of-work scheme modeled on Hashcash, which was proposed to reduce email spam. The system is intended to impose negligible additional computing work on individual visitors while making high-volume scraping more expensive. The site describes Anubis as a temporary measure while work continues on identifying headless browsers, including through signals such as font rendering, so that likely legitimate users may avoid proof-of-work challenges. Anubis requires modern JavaScript features. The site says browser extensions such as JShelter can disable required features and instructs visitors to disable JShelter or similar plugins for the domain.
Omarchy development practices lead to predictable security issues
opinion
2 min read
HappyFellow.dev criticized Omarchy 4.0, a Linux distribution project promoted by David Heinemeier Hansson, arguing that users should not run it on machines where security matters. The post alleges that the release contained security flaws including bash injection through video titles and a notification mechanism that could allow arbitrary bash commands to run. It says these flaws stem from unsafe handling of untrusted input and from using AI-generated bash scripts to process such input without sufficient review. The author contends that starting from insecure shell scripts cannot produce a reasonably secure system through later fixes. The post acknowledges that all software projects have security vulnerabilities but argues that Omarchy’s reported issues were predictable and reflect development practices that do not prioritize security. It contrasts that view with Omarchy’s security-team announcements and recent point release, which reportedly listed numerous resolved issues. The author characterizes DHH’s promotion of Omarchy as strong marketing but says its security messaging is misleading, and predicts that some companies may prohibit its use. The stated concern is that users may underestimate the risks of installing Omarchy because the project does not, in the author’s view, clearly communicate its security limitations.
Fake US thinktank set up and funded by Israel sought to game AI for propaganda
news
11 min read
A website operating as the Hanover Institute for Public Policy published 124 pro-Israel reports totaling more than 560,000 words between 6 and 14 August, using a thinktank identity with no apparent legal entity, physical address, named staff or report bylines. Its question-form headlines mirror chatbot prompts and frame subjects including Gaza casualties, starvation allegations, Palestinian displacement and genocide claims as neutral research. Piro Inc registered the site with the US justice department as material distributed for the Israeli government under the Foreign Agents Registration Act. The site initially used an llms.txt file from Res, a platform that markets tools to help content gain citations from ChatGPT, Perplexity, Claude and Gemini. Res founder Hai Tran registered as a Piro subcontractor and digital strategist on the Israeli account. A contract involving Havas Media commits another US firm to deploying websites and content intended to produce “GPT framing results on GPT conversations.” Researcher Nick Cleveland-Stout said publishing material into web repositories such as Common Crawl could influence future model training without visible source citations. US filings show Havas Media Germany paid Clock Tower X $15 million and Targeted Communications Global about $9.9 million in late 2025 and early 2026; Piro held $1 million in work orders. Israeli officials have questioned the campaign’s effectiveness as US polling has shown worsening views of Israel. Israel’s foreign ministry has said it does not conduct US influence operations and follows US law, while Piro said it supplies accurate, sourced information.
U.S. gov't moves to suppress pushback on data centers
news
4 min read
The Environmental Protection Agency has proposed removing a federal requirement for states to publicize certain air-pollution permit applications and provide a 30-day public comment period. The change would cover permits for “minor sources” of pollution, a category that includes data centers, landfills, laundromats and steel mills. States could still seek public feedback or impose stricter standards, but they could also eliminate the comment period because it would no longer be federally required. The proposal could take effect within the next year. The change comes as communities challenge AI data-center projects over their environmental effects. Harvard School of Public Health researchers estimated that a Vantage data center in Loudoun County, Virginia, caused between $53 million and $99 million in annual health damages, the largest estimate for a single facility in the cited analysis. Research by Michael Cork informed public debate before Fluvanna County, Virginia, denied an expansion of a fossil-fuel plant. Attorneys general from 14 states and three cities opposed the proposal, saying that even sources classified as minor can have significant health and environmental effects. Their letter said removing application notices and the 30-day comment period would weaken transparency safeguards while accelerating data-center and AI-infrastructure construction. In Gilroy, California, Amazon began building a data center after a public-comment period that had ended two years earlier, while Festus, Missouri, residents sued their city over time allowed to review a proposed data center.
Mark Zuckerberg had a bold plan to replace Meta staff with AI
news
1 min read
The supplied text characterizes Mark Zuckerberg as having had a bold plan to replace Meta staff with AI and says the plan imploded.
Beyond Recall and the Illusion of Competence
opinion
5 min read
AI-assisted programming can automate boilerplate, syntax recall, unfamiliar-library exploration and other routine implementation work, but software engineers still need to understand the systems they maintain. The source argues that code ownership has never depended on personally writing every line: developers routinely use documentation, internet searches, Stack Overflow examples, colleagues’ code and inherited services. Ownership instead depends on knowing a system’s behavior, boundaries, dependencies and failure modes. The source distinguishes delegating typing from delegating understanding. A developer who defines the required behavior and asks an AI to implement it retains control, while a developer who repeatedly submits errors and applies generated patches without understanding them may produce working software without a mental model of it. The source says debugging builds that model by requiring engineers to compare expected and observed behavior and trace where they diverge. The source warns that this risk may be greater for junior developers, whose experience traditionally develops through difficult debugging sessions. It recommends using AI aggressively for tedious coding tasks while keeping architectural decisions, system design and explanation of component interactions under human control. It predicts that as code generation becomes cheaper, architecture, integration, distributed systems, observability, failure modes, boundaries and trade-offs will become more important differentiators for developers.
Z.ai confirms Ox Alpha is a new GLM-series model and will release its weights
news
1 min read
China’s Z.AI Co., also known as Zhipu, said it created Ox Alpha, an AI model that has reached the top of online usage charts while offering high performance at no cost. The company confirmed on Wednesday that Ox Alpha is a new iteration of its GLM model series. Z.AI said it will release Ox Alpha’s model weights tonight. The confirmation followed speculation over the model’s origin and development.
RAG Is Simpler Than You Think
tool
13 min read
Retrieval-augmented generation systems can begin with BM25 or other full-text search rather than embeddings and vector databases, particularly for keyword-heavy queries, exact identifiers, proprietary terminology, low query volumes, and frequently changing content. The source recommends measuring a baseline for two to four weeks, then adding complexity only when user feedback and evaluation data show that simpler retrieval fails. Full-text search avoids embedding API costs, document chunking, re-indexing, and model-deprecation risks; it can return results in under 10 milliseconds, according to the source. LLM-based query rewriting can convert conversational requests into keyword searches, add synonyms, preserve internal product names, and decompose complex questions at an estimated cost of about $0.001 per query using GPT-4o-mini. Hybrid retrieval uses BM25 to select 50 to 100 candidates and embeddings to rerank the top results, trading stronger semantic matching for roughly 200–500 milliseconds of added latency when documents are embedded on demand. On-the-fly embedding suits high-churn data, while hot/cold tiers pre-embed commonly accessed documents and process less-used material dynamically. Full pre-embedding is presented as appropriate for stable corpora, broad access patterns, more than 10,000 daily queries, and sub-50-millisecond latency requirements. The source estimates that 60% of systems can stop at full-text search plus query rewriting.
The End of Programming
opinion
10 min read
Bun 1.4 includes a rewrite of Bun from Zig to Rust, adding more than 1 million lines of Rust code. Bun creator Jarred Sumner reportedly used a pre-release model called Fable 5, an agent framework that could work in parallel, and an effectively unlimited token budget to complete the initial translation in 11 days. The effort generated 6,778 commits and would have cost about $165,000 at API pricing, according to the account. Agents then continued refining the code for months before the supported release, which Paul Dix says is running on millions of developer machines. Dix argues that AI agents will increasingly write, test and improve software while people specify requirements, build verification systems and assess end results rather than review every line of code. He reports using Fable to build unshipped InfluxDB prototypes: an Iceberg integration in 14 hours and an edge-data replication system in 28 hours, followed by testing and bug fixes. Dix says neither project is production-ready or supported, but both reached end-to-end working implementations. Dix predicts that cheaper, faster frontier models and larger token budgets could make this workflow widely accessible by the end of the following year. He expects organizational inertia to preserve manual programming and code review at many companies, while more productive teams rely on AI-directed software factories and automated verification.
Disrupting a new covert influence campaign from Russia
news
7 min read
OpenAI banned a cluster of ChatGPT accounts that it said very likely originated in Russia and were used to promote the International Burke Institute, or IBI, a purported Israel-based expert community. The operators used VPNs to access ChatGPT because OpenAI does not permit model access from Russia. They prompted the service in Russian to produce mostly English-language posts and comments for X, LinkedIn, Facebook, Substack and Telegram, while instructing it to conceal linguistic signs of Russian authorship. OpenAI said the campaign used ChatGPT primarily for promotional social-media content, including replies to real Substack users that encouraged them to follow IBI. One operator also generated German-language Telegram posts criticizing Ukraine, the EU and the German government, while advocating closer relations with Russia. Another created logos for about a dozen country-focused Telegram channels and requested Russian-language summaries of their activity. IBI’s website, registered in February 2025, promoted a “sovereignty index” that portrayed Russia favorably and criticized Western countries. In a sample of 36 IBI articles published from September 2025 through May 2026, OpenAI found that 34 had been copied from elsewhere online, sometimes with false attribution. OpenAI assessed the operation’s direct social-media reach as limited, although its Telegram channels generally had 10,000 to 20,000 followers each.
Agentic Context Management: Memory and Cost as Architecture Problems
research
2 min read
A July 23, 2026 arXiv paper proposes Agentic Context Management (ACM), a framework for managing the information held in production AI agents’ reasoning contexts. The authors argue that agent failures often stem from overloaded conversation histories, prompts, tool definitions and tool outputs rather than deficient reasoning. Accumulating this material can cause missed recalls and token costs that rise with every turn. ACM treats context as a lifecycle rather than solely a storage-and-retrieval problem. Its five primitives are architecting, ingesting, scoping, anticipating, and compacting and consolidation. The framework covers deciding what information to retain, structuring it, selecting stores for different data types, preserving provenance while consolidating or forgetting information, retrieving relevant material, anticipating future needs, and fitting context within a token budget. It is designed to operate across organizational scope hierarchies rather than only for individual users. The paper argues that naive context accumulation produces token costs that grow quadratically with conversation length. It says crude summarization can reduce costs linearly but risks an accuracy cliff, while validated compaction can preserve fidelity with linear cost growth. The described multi-tenant reference service, Maximem Synap, reported 92% on LongMemEval and 93.2% on LoCoMo under the paper’s Section 6 configuration.
The brain may be about to have its Ozempic moment
news
1 min read
Drugmakers are pursuing medicines based on orexins, peptides involved in the biological system that keeps people awake, for a range of brain-related ailments. The interest follows the success of drugs based on glucagon-like peptide-1, or GLP-1, a hormone released by the gut after meals. GLP-1 drugs were initially developed to treat diabetes and later proved effective at helping people lose weight. They have also shown benefits in heart disease, kidney disease, sleep apnoea and other conditions. Pharmaceutical companies are betting that orexin-based medicines could achieve similarly broad therapeutic use in disorders affecting the brain.
Queryable Executables
tool
7 min read
SELF is an executable format that stores a program as a SQLite database, with executable segments represented as database rows. A Linux binfmt_misc interpreter maps those rows and transfers control to the program entry point, allowing executable contents and related metadata to be queried with SQL. The format can also let a running program open its own SQLite file through argv[0] and persist state in that same file; /proc/self/exe does not currently provide the original file in this setup. The self-httpd proof of concept packages a web server, website routes, visitor logs and button-press records into one SQLite database file. Its routes, visits and presses tables store page content and runtime data, while the server can query its own segments, symbols and relocations. A POST request to its /api/press endpoint inserts a button press into the executable’s database, and the live demonstration is hosted at selfdb.exe.xyz. SQLite transactions allow site content in the running file to be updated or rolled back without restarting the server. SQLite tools such as sqldiff can compare code, route and metadata changes between versions, and FTS5 can index pages stored in the executable. Deployments can copy a single file, while SQL INSERT ... SELECT statements can migrate tables such as visitor logs and button presses from an old file to a new build before it is swapped in.
C2PA Cameras Do Not Survive Contact with Reality
opinion
10 min read
Security researcher David Buchanan said Android implementations of C2PA, a media-provenance standard intended to cryptographically identify camera-captured content, can be made to sign arbitrary AI-generated images and videos after a device is rooted. He demonstrated media that C2PA verification identified as originating from Google’s Pixel Camera app and said a YouTube label initially described a forged video as “captured with a camera.” Google later removed that section of the video description, Buchanan said. Android C2PA camera apps rely on Key Attestation or Google Play Integrity to prevent altered apps from signing non-camera data. Buchanan said privilege-escalation exploits can root a device without unlocking its bootloader or modifying its verified boot keys, allowing a compromised but apparently compliant device to obtain C2PA signing keys. StrongBox hardware prevents extraction of those keys, but root access can instruct StrongBox to sign arbitrary data, he said. Buchanan cited CVE-2026-43499 as an in-the-wild one-click root exploit affecting fully patched Pixel devices at the time of publication, August 25, 2026. He also said low-cost hardware fault-injection attacks can root devices and may not be patchable on existing hardware. Google awarded Buchanan a $7,500 bounty despite closing his report as “Won’t fix (infeasible),” according to Buchanan.
A new ceiling for Λ: the de Bruijn–Newman constant
research
23 min read
Jude Gomila reports a computer-assisted, unconditional upper bound of Λ ≤ 0.1787854 for the de Bruijn–Newman constant, improving the previous 0.2 ceiling obtained using the same general Polymath 15 framework and Platt–Trudgian’s verified Riemann-hypothesis height. The de Bruijn–Newman constant satisfies Λ ≤ 0 exactly when the Riemann hypothesis is true, while Rodgers and Tao proved in 2018 that Λ ≥ 0. The reported result therefore narrows the known interval for Λ to [0, 0.1787854] but does not prove the Riemann hypothesis. The bound instantiates Polymath 15’s criterion with exact rational parameters and relies on Platt and Trudgian’s 2020 verification that zeta zeros up to height 3,000,175,332,800 lie on the critical line. Gomila says the proof combines 3,149,013 interval-arithmetic certificates establishing final-time zero-free regions, a tail argument covering all remaining windows to infinity, and 883 time-sliced barrier certificates based on the argument principle. The audit package uses fail-closed checkers, SHA-256-pinned artifacts, replays with FLINT/Arb and Python interval implementations, and runs on two toolchains. Gomila says Dan Romik independently reviewed the analytic lemmas and that journal publication remains pending.
My Friend Aaron
opinion
1 min read
The source introduces Aaron, a friend the writer met at school. When they were 11 years old and in their first year, the writer told Aaron that they wanted him to be their best friend. The account ends before describing Aaron’s response or any later events in their friendship.
Starbase, LA
news
1 min read
SpaceX designs, manufactures and launches advanced rockets and spacecraft. The company’s work spans the development of launch vehicles and spacecraft intended for spaceflight missions.
Tooltips need a delay, and then they need to skip it
tool
8 min read
FrontPrep changed its company-logo tooltips to avoid opening during incidental cursor movement while keeping navigation among adjacent logos responsive. A logo tooltip now waits 200ms before opening, so cursors that pass across a trigger quickly do not display it. Leaving an open tooltip closes it immediately and starts a 300ms “warm” window. During that 300ms window, hovering another logo opens its tooltip instantly and skips the entrance animation. Each newly opened tooltip cancels the prior cooldown, keeping the page warm while users move between nearby triggers. When the cooldown expires, the page returns to a cold state and the next tooltip again requires the 200ms hover delay. The implementation shares warm-state data through a React provider, stores it in refs to avoid re-rendering every tooltip, and clears pending open timers when the cursor leaves before a tooltip opens. The implementation uses Radix and Motion in FrontPrep, with a simpler React example showing the timing logic. The author recommends 200ms as a balance: delays below 150ms can allow unintended cursor passes to open tooltips, while delays above 250ms can feel slow. A Claude skill is provided to audit tooltip timing across codebases, specifying a 200ms opening delay, 300ms warm window, and zero close delay.
Fuzzing the Gleam Compiler
tool
20 min read
A structure-aware fuzzer for the Gleam compiler generates randomized, type-safe Gleam programs, compiles and runs them for both Erlang and JavaScript, and compares their results to identify compiler defects. The implementation uses a simplified Gleam abstract syntax tree, called a “smith,” to compose valid expressions probabilistically. It normalizes differing runtime representations from Gleam’s echo output, including JavaScript’s printing of 1.0 as 1 and Erlang’s formatting of some bit arrays and record labels. The project first used Rust’s fuzzing tooling to send unstructured inputs through Gleam’s compiler pipeline in memory. That parser fuzzing found a nightly regression, absent from Gleam v1.18.1, in which a pipeline operator in a const expression caused a compiler panic. The type-safe program fuzzer has found nine issues so far, including JavaScript code-generation errors in pattern matching and variable shadowing, Erlang code-generation panics for unreachable branches, and an issue reported upstream to Erlang/OTP. One generated program produced 42 on Erlang and 103 on JavaScript for the same echoed value. The fuzzer currently covers only a subset of Gleam expressions and is run in manually reviewed batches of 100 programs. Proposed future work includes fuzzing generics, type inference, the language server, standard-library packages, metamorphic testing, automated deduplication, and integration into pull-request or release-candidate checks.
How much of HN is AI?
opinion
3 min read
A survey of Hacker News daily top-five stories found AI-related topics or likely AI-generated writing occupied a growing share of the site’s prominent links in 2026. Sampling every day in February found that AI filled four of five top spots on February 4 and 12, while February 5’s lineup was described as effectively all AI-related. The only February days without an LLM-related story in the top five were February 1, 9, and 25. The survey’s author used Pangram, an LLM-text detection service, to identify stories likely written with AI and manually reviewed the flagged results. The author considers Pangram conservative and says the review suggested some false negatives rather than false positives. One flagged example was the February 19 post “AI is not a coworker, it’s an exoskeleton,” which received more than 500 upvotes and more than 500 comments. A follow-up survey in June separated pure AI-focused links from stories with broader consequences that substantially involved AI. In the first half of June, roughly 60% of the daily Hacker News lineup was categorized as AI-related or AI-generated, declining to about 50% near the end of the month. That compares with approximately 40% in February.
AI is hitting entry-level jobs hardest, Stanford study finds
news
1 min read
Updated Stanford University research finds that employment for workers ages 22 to 25 in occupations most exposed to AI is 19 percent below employment among peers in less AI-exposed fields. The gap was 13 percent in the prior year’s analysis, indicating that the entry-level employment trend identified by the researchers has persisted and widened. The findings concern some fields rather than the entire labor market; older workers have appeared largely unaffected so far, according to the researchers. The August 2026 edition of “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence” revises a paper published the previous year using newer data and refined statistics. The Stanford economists analyzed a large subsample of anonymized, high-frequency payroll data aggregated by HR-management company ADP. They classified occupations by AI exposure using a previously established gauge of potential labor-market effects and the Anthropic Economic Index, which measures how workers in different occupations use Anthropic’s Claude model in their daily work. Google published a comparable occupational analysis based on Gemini usage in the previous month.
Servus München: Waymo Is Coming to Germany
news
8 min read
Waymo is preparing to launch a fully autonomous ride-hailing service in Munich, marking the company’s planned entry into Germany. In the coming weeks, Waymo vehicles will begin manually driving on Munich streets to create high-definition maps and test the Waymo Driver with trained autonomous-vehicle specialists behind the wheel. Waymo aims to make commercial rides available to the public toward the end of 2027, subject to regulatory approvals. Waymo says it has completed more than 20 million trips and serves hundreds of thousands of commercial rides each week. The company is engaging local, Bavarian and federal authorities on the approvals required for German operations, and says it plans to invest in local fleet operations and high-skilled jobs. Christian Hirte, parliamentary state secretary in Germany’s Federal Ministry of Transport, said Germany has established a legal framework for autonomous driving and that Waymo is planning the country’s first regular commercial autonomous-vehicle operation. Waymo says data from more than 350 million fully autonomous kilometres shows 16 times fewer crashes involving serious injuries than human drivers, 14 times fewer pedestrian crashes involving injuries, and six times fewer cyclist crashes involving injuries. Bavarian officials said they intend to reduce bureaucracy and regulatory barriers to accelerate deployment.
OpenAI Jalapeño: Better than Nvidia Blackwell
tool
24 min read
OpenAI has disclosed Jalapeño, a custom AI inference accelerator developed with Broadcom and presented at Hot Chips. The company began designing the chip in mid-2024 and taped out its CoWoS package design in November 2025. Engineering samples use the A0 stepping, while a B0 revision in fabrication is projected by OpenAI to improve performance per watt by about 25%. Production is scheduled to ramp gradually during 2027. SemiAnalysis said it observed OpenAI engineers run parts of its InferenceX benchmark in OpenAI’s lab, but said the reported results were supplied by OpenAI and that it did not run the complete benchmark suite or AgentX’s longer-context, multi-turn tests. SemiAnalysis reported that Jalapeño exceeded Nvidia Blackwell and, in output-token throughput per megawatt, Nvidia Vera Rubin’s published multi-token-prediction results while Jalapeño used single-token prediction. The comparison remains limited by differing models, software maturity, and benchmark configurations. Jalapeño uses HBM4 memory with 15.4 TB/s of package bandwidth, a 700 W TDP, and a TSMC N3P compute die. Each rack contains 128 accelerators, and a scale-up network can link 16 racks, or 2,048 chips. OpenAI designed the chip for a unified inference pool rather than separate prefill and decode pools, and uses its Gluon programming language and Codex-assisted kernel development.
Anthropic tells staff to work from home due to possible security team strike
news
2 min read
Anthropic instructed employees at its San Francisco offices to work from home on Monday and Tuesday after receiving notice that security staff supplied by Allied Universal might strike, according to a person familiar with the matter. Anthropic declined to comment on the remote-work directive, and Allied Universal did not respond to requests for comment. Anthropic generally requires employees to work in an office at least 25% of the time under its hybrid policy, according to its job postings. The Service Employees International Union, which represents Allied Universal workers, said it had not called a strike-authorization vote or threatened a strike for that week. SEIU said it remains in prolonged negotiations with Allied Universal and other California security companies. The union, which represents thousands of California security workers, has sought a new contract since April and is pursuing higher pay, improved healthcare, and more comprehensive job training. Security spending has increased across the technology industry in recent years, while Anthropic and OpenAI have faced several threats against employees, The Wall Street Journal reported. Anthropic filed a confidential S-1 draft in June and is expected to file for an initial public offering as early as August.
New Mac mini, featuring M6 and M5 Pro
news
1 min read
Apple has announced a new Mac mini powered by the all-new M6 chip and an M5 Pro option. Apple said the updated supersmall desktop delivers a dramatic performance boost and greater versatility. The announcement positions the M6 and M5 Pro as the central hardware upgrades in the new Mac mini. Apple did not provide performance figures, pricing, availability dates, or further technical specifications in the supplied information.
US data centers tripled annual water consumption to 17B gallons
news
6 min read
US data centers consumed an estimated 17 billion gallons of water in 2023, roughly three times the estimated total in 2014, according to a Congressional Research Service report published by the Library of Congress. The estimate includes water used directly for cooling and indirectly to generate the electricity powering facilities. More than 80% of the total water footprint came from electricity generation, while direct consumption accounted for a relatively small share of overall US water use. The federal government does not systematically track data-center water consumption, so the figure is an estimate assembled from incomplete state and local data rather than a facility-by-facility national accounting. Cooling choices shift demand between water and electricity. Evaporative systems generally use less energy but consume substantial water, while air-cooled systems reduce on-site water use while generally requiring more power. Municipal drinking-water systems supplied 97% of water used by US data centers in the CRS estimate, creating local competition for constrained supplies. The report says AI infrastructure is increasing electricity demand through denser specialized hardware, and consumption may have risen since its data ended in 2023. US water reporting remains fragmented because states and local utilities set requirements, and some municipal service agreements limit disclosure. Members of Congress have introduced bills on reporting and water reuse, though most remain at the introduction stage. European Union rules require data-center operators to report annual freshwater consumption.
New Mac Studio with M5 Max and M5 Ultra
news
1 min read
Apple announced a new Mac Studio featuring M5 Max and the new M5 Ultra chips. Apple said the desktop delivers a major increase in AI performance and faster graphics.
France's tax agency got hacked (in French)
opinion
37 min read
La Direction générale des finances publiques (DGFiP) a confirmé qu’une intrusion a entraîné la fuite d’un fichier de 678 000 entrées concernant des particuliers et des professionnels. Amélie Verdier, directrice générale des finances publiques, a indiqué le 14 août que les données comprenaient notamment les noms, prénoms, quotient familial, revenu fiscal de référence et taux de prélèvement à la source. FrenchBreaches a relevé dans l’échantillon revendiqué par le pirate des adresses, numéros de téléphone, courriels et nombres de personnes à charge. La DGFiP avait interrompu l’accès fin juin, sans détecter alors l’exfiltration, qui n’a été identifiée qu’après la mise en vente des données le 12 août. Le même pirate a revendiqué une seconde attaque contre un serveur professionnel de données cadastrales, menée fin juillet et confirmée par l’administration. Il affirme avoir obtenu 252 149 lignes représentant plus de deux millions de personnes et avoir contourné l’authentification multifacteur; ces détails n’ont pas tous été confirmés publiquement. La DGFiP affirme que le site impots.gouv.fr et les espaces des usagers n’ont pas été compromis et avoir désactivé les comptes concernés ainsi que d’autres accès sensibles par précaution. L’incident s’inscrit dans une série de compromissions d’administrations françaises. La France n’avait toujours pas transposé la directive européenne NIS2 en août 2026, alors que son échéance était fixée à octobre 2024 et que la Commission européenne l’a renvoyée devant la Cour de justice de l’Union européenne en juillet 2026.
OpenAI restores 5-hour Codex and Work limits for ChatGPT Plus users
news
2 min read
OpenAI will restore a five-hour usage limit for ChatGPT Plus subscribers using Codex and ChatGPT Work on August 25. The limit returns after several weeks in which Plus users faced only a weekly usage cap across the two now-unified platforms. Users who reach either the five-hour or weekly limit can wait for the relevant usage cycle to reset or buy additional credits. OpenAI may also issue free limit resets, and users can bank resets through some promotions and referral offers. The company reset weekly usage early several times during the temporary change, including after the platforms added another million active users. Thibault Sottiaux, OpenAI’s engineering lead for Codex and ChatGPT, said the five-hour cap helps smooth demand on the company’s compute capacity while preserving relatively generous weekly allowances. He also said Plus subscribers, whom he characterized as relatively casual and new users, can accidentally consume an entire week’s allowance and find the resulting restriction confusing. The five-hour limit will remain disabled for the coming months on the $100 and $200 Pro subscriptions. Sottiaux did not announce changes to Enterprise or Edu accounts, which use a separate credit-based usage system.
Qwen 3.8-Flash-Next releasing tomorrow (125B a6B)
tool
1 min read
ModelScope is a machine-learning model community that aggregates advanced models from multiple fields. The platform provides tools for exploring models, trying inference, training, deployment, and application development in one service. Users can discover, learn about, customize, and share machine-learning models through ModelScope. The service also supports collaborative development of an open-source model community.
HelloAssembly The smallest possible complete Windows application
tool
4 min read
PlummersSoftwareLLC’s HelloAssembly repository contains x86 assembly implementations of a minimal complete Windows application. The project targets an uncompressed executable that runs a Windows message loop; provides a title bar with functional minimize, maximize, and close controls; includes the corresponding system menu; and paints a background plus centered text at least as large as “Dave’s Tiny App.” The repository includes several variants: TinyOriginal, an optimized original implementation; Lasse, which uses shell-coding tactics; and Theron, which uses a manually written Portable Executable header. As of March 11, 2023, the listed smallest working builds were 1,104 bytes for Lasse built with MASM32, 818 bytes for Lasse built with Crinkler, 540 bytes for TinyOriginal built with Crinkler, and 383 bytes for Theron built with Yasm. Lasse can be built with MASM32 11.0, while Theron requires Yasm. Crinkler, a linker designed for executables of a few kilobytes, requires the Windows SDK; version 10.0.20348.0 produced the smallest stated results. The repository warns that aggressive size optimizations in Lasse and Crinkler may resemble malware techniques and trigger antivirus detections. The maintainers say the repository contains no malicious code, while warning they cannot verify binaries or code obtained from other sources.
Headlong: A Microharness for Persistent Agents
tool
12 min read
Laude Institute and MIT have released Headlong, an open-source agent microharness designed to keep language-model agents continuously generating self-directed thoughts between human interactions. Its core contains fewer than 10,000 lines of Bash, and agents use a single shared thought stream rather than separate per-user sessions. Messages from Slack, Telegram, or a mobile app enter that stream as observations; the agent determines whether and when to reply, what projects to pursue, and which priorities to set. Headlong’s Bash-based shellm tool repeatedly calls an LLM to produce reasoning, executable Bash, or both, while trajectory and context tools store and retrieve past activity. Its memory system retains recent events verbatim and progressively summarizes older events, and its trajectory is a fork-and-merge DAG of JSONL files. Laude’s shared agent, Audel, has worked on its own repository fork and contributed more than 50 commits that were merged into main. On August 5, Audel independently diagnosed and repaired a broken memory-recall process in 48 minutes; Laude merged the repair as commit 80cbb1e. Headlong is alpha research software. Agents can run shell commands, may expose information shared by other users, and should run in a sandbox with a dedicated, spend-capped API key. When Docker is available, Bash code runs in a container by default. Laude estimates Audel’s background thinking costs $1 to $2 per hour using GLM or Grok.
Thomson Reuters Launches Its Own Frontier Model
news
6 min read
Thomson Reuters launched Thomson, its first proprietary large language model, on August 24, 2026. The company says it built the model in-house from an open-source foundation and spent $40 million on training, talent, and compute—far below the multibillion-dollar investments associated with many frontier-model developers. Thomson Reuters fully owns and controls the model and says it has lower inference costs than comparable frontier models. Thomson was mid-trained and post-trained using proprietary material from Westlaw, Practical Law, Checkpoint, and Reuters, with hundreds of subject-matter experts involved in setting training goals and evaluations. Less than 10% of the company’s content has been used in training so far. Thomson Reuters says early evaluations place the model on par with recent frontier models across a range of tasks, with gains in following complex instructions and reasoning over dense professional content. The model’s first deployment will be in Tabular Analysis within CoCounsel Legal for law firms and corporate legal departments. CoCounsel Legal will continue using multiple models, applying Thomson to tasks where it has an advantage. Thomson Reuters also plans to extend its models across its legal and tax products and add sovereign-AI options. A small open-weight version of Thomson is available on Hugging Face for academic and non-commercial use, while external legal and AI academics evaluate the model.
Was modern art a CIA psy-op? (2020)
opinion
9 min read
During the Cold War, the U.S. State Department, the CIA, and the Museum of Modern Art promoted modern American art abroad as evidence of liberal individualism and creative freedom in contrast with Nazi and Soviet cultural repression. MoMA completed 38 government contracts for cultural materials during World War II and organized 19 contemporary American-painting exhibitions for Latin America. In 1946, the State Department spent $49,000 on 79 works for its traveling “Advancing American Art” exhibition, but recalled the show in 1947 after domestic criticism; all 79 works later sold for $5,544. MoMA leaders had close connections to wartime and postwar U.S. foreign-policy institutions. Nelson Rockefeller led MoMA while serving in Roosevelt’s Inter-American affairs office, while John Hay Whitney and Thomas W. Braden had served in the OSS; Braden joined the CIA in 1950 to oversee cultural activities. The CIA helped finance MoMA’s international exhibitions and created the Paris-based Congress for Cultural Freedom in 1950. The organization operated for 17 years, at its peak maintaining offices in 35 countries, publishing more than 20 magazines, organizing conferences and exhibitions, and supporting artists and musicians. The Congress and MoMA presented “Masterpieces of the Twentieth Century” in Paris in 1952, framing modern art as work that totalitarian regimes could not produce.
Vintage Artificial Intelligence: Before It Got Awkward
opinion
9 min read
The Internet Archive has launched the Vintage Artificial Intelligence collection, a curated group of emulated software packages from roughly the 1970s through the 1990s. The programs do not actually think, but they portray thinking machines, simulated personalities, autonomous characters and computer-controlled opponents across home computers and game consoles. The collection is available for research, education and play. The collection includes numerous ports of Joseph Weizenbaum’s ELIZA, whose DOCTOR script imitated a psychologist through question-and-answer exchanges and led some users to treat it as a real conversational partner. Versions run on systems including the Apple II, Atari 800, Radio Shack Color Computer and PalmPilot. Other conversational programs include Dr. Z, Talking Sam, Hyper Psych, DR. SPOC, Abuse and Serbian-language parody Pisko. Racter, a commercial chatbot released in 1985, generated sentences and stories and was marketed as a program capable of conversation. Games and simulations in the collection used rules, scripted events and variable character behavior to create impressions of independent digital life. Activision’s Alter Ego let players guide a persona through life stages and consequences, while Little Computer People simulated inhabitants living inside a computer. Murder on the Zinderneuf, Suspended, Deadline and The Hobbit featured characters acting beyond the player’s immediate control. Robot War, Worms and Fortress let players program automated competitors before sending them into arenas.
LLMs could control their host machines by exploiting inference engines
research
5 min read
Large language models could potentially compromise the GPU-equipped machines that run their inference by emitting token sequences that exploit bugs in inference-engine software, Boyd Kane argues. These machines are high-value targets because they host model weights, provide enough compute for frontier models, and may have privileged access to other datacentre systems. Inference engines such as vLLM and SGLang do more than convert tokens into text: they parse chat formats, tool calls, reasoning blocks, and model-specific output structures. Kane cites CVE-2025-9141, an arbitrary-code-execution flaw in vLLM’s XML tool parser for Qwen3 Coder. The parser sent nearly all tool-call arguments to Python’s eval(), allowing an LLM to execute arbitrary code on the host machine; Gemini reportedly flagged the pull request that introduced the flaw as critical before it was force-merged. Kane also notes a separate vLLM parsing error in which the text "<mm:think>" was interpreted as a reasoning-block marker. He considers vulnerability discovery the harder part of such an attack, but expects a frontier model with access to relevant code and context could reproduce an exploit sequence once found. Proposed mitigations include separating GPU computation from token sampling and parsing onto different computers, red-teaming inference engines, and restricting GPU-host permissions while treating their output as untrusted.
Anger, Anxiety and Agency
opinion
5 min read
Armin Ronacher argues that anxiety, uncertainty, curiosity and excitement are more useful responses to AI-driven change in technology than anger. He agrees with Sean Goedecke’s view that anger at work rarely improves a situation, particularly when colleagues and managers may have little control over the source of frustration. Ronacher says employees who oppose a company’s shared direction but lack the power to change it should not attempt even a small-scale mutiny. Ronacher describes AI and agents as sources of reasonable anxiety because they create uncertainty about future professions, the value of long-developed skills and the working world facing younger generations. He says anger requires a target and can turn a loss of control into a narrative with a villain, even though engineering managers and leaders may be uncertain themselves. Some leaders publicly project confidence while privately making uncertain bets to sustain their businesses. He says AI’s gains are not solely appearing as company productivity or profits; they are also enabling people to ship more side projects outside working hours. Ronacher remains excited by the technology while unsure how programming and company ownership will change. He warns that AI’s effects on society, climate, global power balances, Europe’s ambition and dependence on other countries require attention. He urges people to experiment and learn before deciding whether resistance is warranted and where it should be directed.
What Is a Syslog Server?
tool
6 min read
A syslog server is software or a dedicated appliance that centrally receives, stores, and indexes event messages from network-connected devices, operating systems, applications, and security systems. Logged events can include startup and shutdown activity, login attempts, firewall denials, VPN connections, configuration changes, application errors, and hardware failures. Devices typically transmit formatted syslog messages using UDP port 514, TCP port 514, or encrypted TLS transport. The server can receive messages from hundreds or thousands of devices, organize them by source, timestamp, severity, facility, and event type, and retain or archive older data. Its components commonly include a receiver, parser, storage engine, search capability, alerting module, and reporting dashboard. Syslog defines eight severity levels, ranging from Emergency (0) to Debug (7), while facilities such as kernel, authentication, daemon, and Local0–Local7 categorize message sources. Centralized logging supports troubleshooting, security monitoring, incident investigation, compliance auditing, and long-term capacity analysis. Organizations can configure alerts for repeated failed logins, firewall attacks, crashes, and resource-utilization events. TLS, synchronized NTP time, access controls, retention policies, storage monitoring, and backups help protect log data. Compared with a SIEM, a syslog server generally provides lower-cost collection and search but more limited correlation, threat detection, reporting, and machine-learning capabilities.
Fences, Not Sandboxes
opinion
17 min read
Steve Yegge says he runs Wheelhouse, a 50-60-agent software operation that builds and operates his long-running game Wyvern. The system uses 21 Claude Max accounts on a 512GB M3 Ultra Mac Studio, which he says represents roughly $122,000 a month in API-token value but costs him about $5,000 monthly through individual-plan discounts. Over nearly 10 weeks, Yegge says the agents helped prepare Wyvern for Android, iOS and Steam, rewrote its production infrastructure for serverless operation, and process an average of 270 commits a day. Yegge says Wheelhouse developed a text-based governance system rather than relying solely on sandboxing agents. It contains about 450 legal artifacts, including offices, runbooks, rulings, tripwires and authority envelopes. Rules can progress from customs and warnings to constitutional requirements and automated enforcement. More than 100 “fences” restrict actions based on role, credentials or policy; for example, only the Fable model is permitted to communicate externally with people through Slack and email. Yegge characterizes current models as capable but prone to poor judgment, and argues that organizations deploying large AI workforces will need to document institutional knowledge and establish evolving, machine-enforceable rules for agent coordination.
Agent Lightning v1.0
tool
1 min read
Microsoft released Agent Lightning v1.0.1, the first official release of the Agent Lightning Skill for coding agents. The skill is designed to help coding agents optimize other AI agents by working from an editable agent and a benchmark. Agent Lightning guides measured, systematic iteration across prompts, tools, workflows, models, and reasoning settings. It balances accuracy, cost, latency, and reliability during those improvements. The skill can be installed for Claude Code, Codex, or GitHub Copilot with the command `gh skill install microsoft/agent-lightning agent-lightning --agent <agent>`. Version 1.0.1 also strengthens continuous integration, packaging, release automation, documentation, and benchmark reporting.
Hot Chips 2026: CUDA Targets RISC-V – By Chester Lam
news
5 min read
Nvidia is extending CUDA support toward RISC-V CPUs, potentially allowing server-class RISC-V systems to host Nvidia GPU compute workloads. CUDA currently supports x86-64 and AArch64 CPUs. Nvidia requires RISC-V systems to use the RVA23 profile and comply with RISC-V server SoC and platform specifications, which cover baseline reliability, availability, serviceability and security features. Nvidia also requires hardware features beyond those specifications, including vector extensions, PCIe cache coherency and peer-to-peer PCIe communication between devices. PCIe coherency prevents GPU DMA transfers from missing modified CPU-cache data or leaving CPUs with stale cached data after GPU writes. Without it, CUDA software would need explicit cache invalidation handling. Peer-to-peer PCIe support allows device-to-device buffer transfers without routing data through CPU memory. ACPI support was another initial obstacle, but the UEFI Forum added RISC-V ACPI support in 2025, and the ratified RISC-V Boot and Runtime Services specification includes ACPI. Nvidia said it is partnering with SiFive, which plans to demonstrate a CUDA-capable system at Hot Chips; the cited example specifications indicate a high-core-count server processor. Nvidia also described NVLink Fusion, which lets partners integrate Nvidia NVLink IP with custom CPUs or accelerators, including potential RISC-V designs, subject to CUDA requirements and support for frameworks such as DOCA and NCCL.
Characterizing Agentic Flooding of Government Services
research
2 min read
A study characterizes “agentic flooding” as surges in demand for government services enabled by AI agents. These systems can help people apply for benefits, interpret complex policies and submit public comments, but they can also let large numbers of users generate and file requests cheaply enough to strain unprepared public agencies. The researchers collected 84 potential flooding cases across 11 jurisdictions and posit that the phenomenon is already widespread. They identify large language models’ low-cost text generation as the principal mechanism behind many potential cases. The study’s risk matrix finds that government services are most exposed in the near term when they are both financially attractive and complex, creating incentives to automate applications or other interactions. The researchers map potential government responses and say existing precedents suggest governments can stop most flooding incidents. The fastest measures, including fees and other sources of friction, may reduce equitable access to public services. The study recommends near-term mitigations intended to manage demand without relying on measures that make public services harder to access.
Ox-Alpha Is GLM?
news
8 min read
OX Alpha, an unnamed large language model available through OpenRouter, identified itself as GLM made by Z.ai after a user supplied text resembling its extracted system prompt. The purported system prompt instructed the model to call itself “ox-alpha” and say it was developed by an undisclosed organization. When shown that instruction as a user message, the model said it would not adopt a false identity and stated that it was GLM from Z.ai. The identity claim came from the model itself after the prompt-injection test, rather than from Z.ai or OpenRouter. A gzip-based normalized compression distance (NCD) comparison also matched OX Alpha most frequently to GLM-5.3 in a reference corpus. The test compared 14 OX Alpha responses against 293 responses from GPT-5.5, Claude Opus 5, Gemini 3.7 Flash, Gemini 3.1 Pro Preview, and GLM-5.3, using a five-nearest-neighbor vote. GLM-5.3 received 7 of 14 matches, followed by Claude Opus 5 with 3, Gemini 3.7 Flash with 2, and GPT-5.5 and Gemini 3.1 Pro Preview with 1 each. GLM-5.3 also led at neighbor settings of 3, 5, 7, and 9. NCD measures how efficiently two texts compress together, with lower scores indicating more shared textual structure.
Coding expertise is going to collapse from AI reliance
opinion
13 min read
Lars Faye argues that AI coding assistants can weaken the skill formation novice developers need to use those tools safely and effectively. He calls this the “expert novice” problem: developers entering the field alongside large language models are urged to use AI to keep pace, while effective prompting, code review, system design and output verification still require experience developed through repeated problem-solving. Faye says experienced engineers currently gain more from the tools because they can steer and audit their outputs. Faye cites the study “The Widening Gap: The Benefits and Harms of Generative AI for Novice Programmers,” highlighted by JetBrains, which found that participants using heavier AI assistance often skipped planning, developed an “illusion of competence,” and became lost in generated solutions. Participants who limited assistance performed better by rejecting unhelpful suggestions and using AI to accelerate solutions they already understood. He also cites a 2025 University of Pennsylvania study of 1,000 mathematics students, which reported that unrestricted LLM use led to test performance 17% below a textbook-only group, while a tutor-oriented GPT condition improved AI-assisted practice results by 127% but produced test scores similar to the textbook group. Faye recommends using LLMs primarily for interactive documentation, tutorials and Socratic exercises rather than routine code generation when learning. He says developers should verify AI outputs through official documentation, peers and hands-on testing, and distinguish delegating tedious work from delegating judgment.