Artificial Intelligence

OpenAI Advances Low-Resource Language Coding: MultiLangBench 87.4%, COBOL Migration Near 98%

On 19 Oct OpenAI said latest models strengthen building and migrating software in low-resource languages such as COBOL, Fortran, Visual Basic, and OCaml. MultiLangBench: GPT‑6 Astra averages 87.4% across 20 languages (vs 26.8% GPT‑5.6 Sol). COBOL migration: GPT‑6.1 Sol 86.6% at low effort / 97.9% at higher effort, costing ~39% less. MultiLangBench releasing on GitHub; production still needs expert review.

October 21, 20264 min read

Artificial Intelligence

Anthropic: Claude Designs Scorpion Neurotoxin Binders; 17 of 211 Pass Dual-Lab Tests

Anthropic 20 Oct 2026: An internal Claude Science model, over ~64 hours and ~10,000 GPU hours, produced 318,484 designs and 211 tested scorpion-neurotoxin minibinders; 17 met dual-lab SPR binder criteria (each vs one toxin only). Not an antivenom; no cellular/in vivo neutralization claimed. All designs and data released.

October 21, 20264 min read

Artificial Intelligence

From the Internal “Soul Doc” to Claude’s Constitution: Anthropic Publishes Its Character Blueprint

On 21–22 Jan 2026 Anthropic published Claude’s Constitution: a values-and-behavior blueprint written primarily for Claude, treated as final authority, CC0, living document. Conflict priority: broadly safe → ethical → Anthropic guidelines → genuinely helpful. Earlier internal nickname “soul doc” confirmed by Amanda Askell as training material; distinct from 2023 Constitutional AI method.

October 19, 20265 min read

Artificial Intelligence

Anthropic’s Public Benefit Fund: 5% of Shares Now, Plus 2% for Every 50% Rise in Value

On 18 Oct Anthropic announced the Anthropic Public Benefit Fund: 5% of shares committed subject to stockholder approval; board may add 2% for every 50% rise in value above a baseline. Wholly owned LLC; no lobbying; combined with founder 80% pledge and employee match ≈20% of shares for public good; $200M already for labor-market impacts.

October 19, 20263 min read

Artificial Intelligence

Anthropic Embeds METR Evaluators: Employee-Like Access for Six, Publication Rights Without Veto

On 15 Oct Anthropic updated independent-evaluator arrangements: six METR staff get employee-like access roughly on par with internal risk teams; METR can publish risk analysis without Anthropic veto (redactions only in predefined categories); three-tier access model; also SecureBio, Accenture, early Hugging Face talks.

October 18, 20264 min read

Artificial Intelligence

HKMA Unveils “ART” Framework: Accountability, Resilience, Trust as Banking AI Moves Toward Agents

On 15 Oct the HKMA unveiled the ART Framework (Accountability, Resilience, Trust): updated AI supervisory guidance, a Cyberport–HKU training-data platform, and GenA.I. Sandbox++ agentic testing. The Finance+A.I. Forum marks end of the second Sandbox cohort; over 60% of use cases explored agentic AI.

October 18, 20263 min read

Artificial Intelligence

AI Drowning Detection and SmartPLAY Anti-Touting: Two LCSD AI Projects Aiming for 2027

Two LCSD AI projects: elevated cameras and computer vision to help lifeguards spot suspected drowning, with a Q1 2027 PoC at Morse Park and Ma On Shan main pools; SmartPLAY uses AI on ~27,000 daily bookings, expected operational in 2027.

October 14, 20264 min read

Artificial Intelligence

Tung Chung Autonomous Vehicle Public Trial Rides: About Two Weeks From 21 October, Yat Tung–Sheraton Loop ~15 Minutes

TD District Council paper: Tung Chung AV public trial rides 21 Oct–3 Nov (14:00–18:00); Yat Tung–Sheraton loop ~5.1 km / 15 minutes; 600 tickets via Islands DC Secretariat.

October 13, 20264 min read

Artificial Intelligence

Gates Proposes Human Reserved: Work Deliberately Kept for People Even When Machines Can Do It

In Gates Notes (2026-08-26), Bill Gates proposes Human Reserved: work kept for people even when AI/robots can do it—nature-reserve analogy, plus economics, dignity, evolution, and token/robot tax.

October 12, 20266 min read

Artificial Intelligence

About 10 Minutes With AI: Study Finds Lower Solve Rates and More Quitting After Help Is Removed

UC Berkeley / COLM 2026 RCTs (N=1,222): after ~10–15 minutes with GPT-5, unassisted fraction and SAT-reading solve rates fell and skipping rose once AI was removed; costs concentrated among direct-answer users.

October 12, 20263 min read