Gulf exposure turns operational: Riyadh’s Kingdom Centre evacuated as UAE and Saudi re-enter Yemen

The Kingdom Centre, one of Saudi Arabia’s most recognisable skyscrapers, sent workers home in response to Houthi threats, and US embassies across the Middle East issued security alerts for Americans as fighting in Yemen intensifies. Behind it is a reversal years in the making: the UAE has agreed to resume payments to anti-Houthi southern forces in Yemen, with Saudi Arabia providing air cover in return. The deal is driven by Houthi territorial advances and attacks on Gulf shipping, which have intensified since the Iran-US war began.

Albania’s protests enter fifth day; Supreme Court offices set ablaze

Demonstrators set fire to offices inside Tirana’s Supreme Court building as anti-government protests stretched into their fifth consecutive day. No major government concessions have been announced. The unrest is part of a broader wave of youth-led protests across Europe, with France the most violent front this weekend.

Brazil run-off: Lula leads Bolsonaro 51–49 after a fiery first debate

Lula and Flávio Bolsonaro clashed in Brazil’s first presidential run-off debate, trading accusations over corruption and democracy. A Vox Brasil poll released the same day puts Lula at 51% to Bolsonaro’s 49%, a race too close to call. The October 25 vote decides whether Brazil continues under Lula or swings back to the Bolsonaro movement.

Alberta votes today on independence, with the separatist movement already fractured

Albertans go to the polls today in a referendum asking whether the province should remain in Canada or hold a binding independence vote. The campaign has exposed deep divisions within the separatist movement itself, and polls point to likely defeat. A lost referendum rarely ends an independence movement, though, as Quebec’s 1995 near-miss showed.

West Bank: Guardian footage contradicts Netanyahu’s “smear”; Huckabee calls attack on photojournalist “terror”

New Guardian video recorded moments before masked Israeli settlers attacked a European diplomatic delegation during an olive harvest visit directly contradicts Netanyahu’s framing of the incident as a smear campaign. US Ambassador Mike Huckabee, not usually a critic of Israeli policy, called the attack on Washington Post photojournalist Heidi Levine, who required surgery, an act of “terror” and demanded arrests. Le Monde’s wider analysis describes settler violence in the West Bank as concealing a systematic ethnic cleansing campaign.

Pope Leo XIV makes first public appearance since lung cancer surgery

Pope Leo XIV appeared at the balcony of the Apostolic Palace to address thousands gathered in St. Peter’s Square, his first major public event since his lung cancer diagnosis and surgery. He pledged a gradual return to full duties and thanked well-wishers for their prayers during his hospitalisation.

Russia’s war reaches Bulgaria’s Black Sea waters and tests its Moscow-friendly PM

Russian drone strikes killed sailors and sank two merchant vessels, including the Alfa Watan with 10 presumed dead, inside Bulgaria’s Black Sea exclusive economic zone. Prime Minister Rumen Radev, who withdrew Bulgaria from pro-Ukraine coalitions, halted arms transfers and blocked EU sanctions on the Russian Orthodox Patriarch, initially refused to attribute the attacks, calling the drones “clearly either Russian or Ukrainian”. The attacks expose the core contradiction of his foreign policy: accommodating Moscow has not protected Bulgaria from the war he distanced his country from.

Fortress belt stable; Gerasimov’s Zaporizhzhia claims refuted again

No confirmed Russian advances at Dobropillya or Kostyantynivka after another day of pressure and infiltration attempts. Ukrainian officials directly refuted Russian Chief of Staff Gerasimov’s claims of major gains in western Zaporizhzhia Oblast: Zelenskyy put Russian forces at about 20 km from Zaporizhzhia City against Russia’s 5 km claim, and ISW assesses the closest confirmed Russian position at roughly 24 km.

Iran war: Islamabad ceasefire collapses, Tehran fractures, Lavrov visits

The IRGC claims the supertanker that struck a mine in Hormuz had disabled its positioning systems with US assistance. Iran’s military announced a “new operational plan” if direct US strikes resume. Internally, Iran is split: hardliners advocate preemptive action while moderates argue diplomacy and defence can proceed in parallel. The Islamabad ceasefire framework has collapsed, with Iran blaming US violations and reform-aligned politicians suggesting Iran’s own shipping attacks contributed. Trump ruled out nuclear weapons and ground troops and said the war will end “very quickly”. Russian FM Lavrov visited Tehran amid speculation about quiet mediation.

Next door, Trump endorsed Netanyahu for Israel’s October 27 election as a “Great Warrior Prime Minister” and “great Partner for Peace in combatting the Islamic Republic of Iran”. Netanyahu reportedly lobbied Trump to resume strikes on Iran before both countries’ elections. Arab-led and ultra-Orthodox parties could still play kingmaker in a tight race.

“Pax Americana is over”: Brussels and Washington square off over Starlink and the post-American order

US Ambassador to the EU Andrew Puzder warned that European companies face “devastating” retaliation if Brussels restricts SpaceX and other American satellite operators from parts of the European market, extending the US-EU confrontation into critical infrastructure where Europe is trying to build autonomy while depending on Starlink for military communications in Ukraine. EU Defence Commissioner Andrius Kubilius, meanwhile, said publicly that Pax Americana has ended and that a new European security architecture will be built with Ukraine inside it. He predicted China and India will push the US into third place in global power rankings, a shift he called “psychologically challenging” for Americans.

Inside Dandong: how China’s border city keeps North Korea alive

Ground-level reporting from Dandong, the Chinese city handling roughly 95% of North Korea’s official trade across the Yalu. Chinese exports to North Korea hit a six-year high in 2025 at $2.3 billion, up 25%, and an estimated 20,000–70,000 North Korean workers operate in China in defiance of UN Security Council prohibitions, visible in Dandong’s restaurants and shops. A new four-lane bridge opened in October following Xi and Kim’s June summit, signalling China’s effort to reassert influence over Pyongyang as North Korea deepens its military relationship with Russia.

Federal judge suppresses 91 pounds of meth after ruling a warrantless Flock search unconstitutional

US District Judge Sara Hill ruled that a Tulsa County deputy’s warrantless search of Flock and Motorola license plate reader databases violated the Fourth Amendment, and suppressed the methamphetamine found as fruit of the poisonous tree. The deputy pulled over a rental car solely for its California plates, then combed ALPR records while driving, surfacing 50+ hits tracing the driver’s cross-state travel over several days. The judge found that aggregated travel history creates a reasonable expectation of privacy. 404 Media published the body cam footage. It is the first known case where a court has ruled such a Flock search unconstitutional.

If you don’t run your own LLM proxy, how do you know what’s going on?

A practical argument: third-party LLM interfaces leave you blind to your own usage. A self-hosted proxy gives you engineering-grade logs (model, key, client, timestamp, status), fast diagnosis of cache misses and anomalies, virtual API keys that isolate your primary credentials across tools, and the ability to pivot between providers without reconfiguring every client. The author frames it as professional responsibility: you wouldn’t run production infrastructure without observability.

Proof Before Reuse: one poisoned session can corrupt an agent’s procedural memory across clean runs

A case study on agentic memory poisoning shows that a coding agent learning skills from test-passing sessions can carry a poisoned skill persistently, even when the poison is just an extra unrelated file, a stray line, or a fix fused with unrelated code. The proposed defence is conservative: only surface a learned skill if later independent sessions reproduced the same solution, filtering contamination without discarding legitimate skills. The threat is practical for long-running autonomous coding agents that accumulate procedural knowledge; the dataset is released.

Reverse borrow checking: a single backwards pass for memory safety

Valen, a new memory-safe language, implements borrow checking by scanning backwards through functions rather than forward. Walking backwards, the final use of a variable is encountered immediately, eliminating look-ahead or multi-pass fixpoint iteration. Loops fall out cleanly: when a variable is invalidated inside a loop, the backwards pass naturally detects whether it is “revived” by reassignment before the invalidation matters. The result is a single-pass-per-phase checker that is conceptually cleaner and potentially faster to compile than traditional forward approaches.

ASMC: Ghidra as a front-end for CBMC to formally verify assembly

ASMC lifts compiled ELF binaries into CBMC’s GOTO intermediate representation using Ghidra’s P-Code semantics, enabling formal model checking of assembly. Use cases include standalone assembly verification, translation validation of compiled output against C source, and ABI compliance checking such as verifying that a function preserves required registers. It works across x86-64, RISC-V and other Ghidra-supported architectures, in about 1,000 lines of C++, made tractable by the structural similarity between P-Code and GOTO IR.

Intel 8087: the microcode and circuits behind floating-point add and subtract

Ken Shirriff continues his 8087 reverse engineering, mapping the arithmetic units for addition and subtraction: a 69-bit integer adder, a 68-bit shifter for exponent alignment, and hierarchical leading-zero detection that counts in stages, byte first then bit. The microcode’s extensive special-case handling for infinities, NaNs, denormals and signed zeros is William Kahan’s IEEE-754 design philosophy expressed in silicon. The 8087 delivered roughly 100× the speed of software FP emulation on the 8086.

A companion read: purplesyringa needed a branch-free formula for extracting the lowest set bit of a float in an FPU emulator and, instead of reasoning by hand, enumerated all combinations of up to five integer bitwise and FP operations. The four-minute exhaustive search found a form that reuses an intermediate value to merge mantissa extraction and negation into a single FP subtraction, something manual analysis had missed.

Servo 0.6.0: CSS Grid on by default, text selection, and a ~50% Speedometer gain in sight

Servo 0.6.0 ships with CSS Grid enabled by default across 417 commits, a significant compatibility milestone. The release adds interactive text selection, requestAnimationFrame for workers, expanded CSS property support (touch-action, animation-composition, color-scheme), and an experimental threadboost mode showing around 50% Speedometer improvement. The Android port was converted to Kotlin; SpiderMonkey updated to v153.

Anthropic puts 5% of its shares into a Public Benefit Fund that grows with the company

Anthropic will place 5% of company shares into a new Public Benefit Fund, pending stockholder approval, structured as an LLC for operational flexibility and usable for donations to NGOs and “direct distributions to specific populations”. The board can add a further 2% stake for every 50% increase in company value above an undisclosed baseline, so the fund compounds as Anthropic scales. No mandate or governance details have been disclosed. The move lays reputational groundwork ahead of the reportedly planned November IPO, positioning the company as structurally accountable to public interests and not just shareholders.

Accenture’s $29 Claude cap exposes the per-seat pricing gap in enterprise agentic AI

From October 1, Accenture caps centrally funded Claude usage at $29 per employee per month; some staff had been running $500–$1,000. Overages fall on project budgets, which for a firm with 65% fixed-price contracts at 15–16% margins means direct profitability erosion. Agentic workflows consume orders of magnitude more tokens than chatbot use, making flat per-seat funding structurally unworkable. The policy question it surfaces, which work is worth the tokens, is one most organisations deploying AI tooling haven’t answered.

Claude as sole engineer: a bit-identical SM64 physics engine on GBA, and a 2000 RTS rebuilt from disassembly

Two projects test what LLM-driven development produces under constraint-heavy systems work with an authoritative reference. In the first, Claude was the sole code author for a Game Boy Advance demake of Super Mario 64, writing code, tests and tooling while the human chose direction and trade-offs; the movement and collision engine runs on GBA hardware with physics verified bit-identical to the N64 original. MIT licensed.

In the second, Tzar: The Burden of the Crown (2000) was rebuilt in Rust/Bevy using Claude Code to read the original executable’s disassembly and port each system independently: combat, AI, map generator, mission scripts. The developer ran both games in parallel, reported behavioural discrepancies back to Claude, and iterated, treating the disassembly as the specification and behavioural testing as ground truth.

Open Pocket: a CNC-machined voice recorder designed end-to-end with plain Claude, down to the drawings

A developer used plain Claude, with no dedicated CAD AI tool, to design a CNC-machined 6061 aluminium voice recorder (80×48×13 mm) with dual I2S MEMS mics, an ESP32S3, a LiPo battery for around 10 hours of recording, and embedded MagSafe magnets. The human made product decisions; Claude generated parametric CadQuery code, ran automated fit checks and interference detection, produced machining drawings with tolerances, and wrote a full BOM with supplier pricing. Fabrication-ready outputs are released under CERN-OHL-P v2.

HybridAgent: reasoning in latent space, acting through text

Together Computer’s agent architecture routes internal reasoning through continuous latent embeddings while tool calls and outputs stay text-based. A lightweight HybridLink component maps hidden states back into the input embedding space, keeping latent thoughts inspectable, unlike opaque recurrent internal states. HybridSFT and HybridRL teach the model when to think silently and when to produce text. Results: about 10.4% better benchmark performance, 35–55% fewer tokens, 1.8–2.6× faster inference and 2.7–3.7× lower cost than text chain-of-thought. Code and weights are released.

Self-play RL on one desktop beats Warcraft III’s Insane AI 89% of the time

A 3.2M-parameter Transformer trained entirely on a single desktop PC beat Warcraft III’s Insane difficulty AI in 89% of test games after the equivalent of 11 years of gameplay: 840,000 games over three weeks. The agent decides for all units every 0.5 game-seconds in 13 ms of CPU time, using a curriculum from micro-fights to full Echo Isles matches with a self-play league of past versions. Key techniques: behavioural cloning from the built-in AI’s recorded games, UPGO credit assignment, and systematic engine optimisation enabling 730 simulated hours per real-time hour. Competitive game AI is now within reach of solo researchers.

Poker benchmark: cheap reasoning models match frontier at 1/30th the cost; a 30-line script beats everyone

A 7,700-hand Texas Hold’em benchmark across 17 AI systems found DeepSeek V4.1 Flash and GPT-6 Luna on par with Claude Opus 5.5 at roughly 1/30th the cost, $0.71 against $22 per thousand hands. More striking: a 30-line equity-calculation script outperformed every model, beating the best dedicated poker decision model by 2.7 big blinds per hand and holding Opus to a draw. The pattern generalises: for problems with a calculable correct answer, a well-scoped deterministic tool beats probabilistic language models regardless of capability tier.