ActuIA
The Artificial Analysis v5 ranking, published 27 October 2026, weights 60% of its score to private or secret-response evaluations; twelve models with published weights lose 12–16 points and Mistral Large 4 falls to 25 points while Claude Opus 5.5 remains at 58.
WWWhat's New
Wwwhatsnew reports that OpenAI released GPT-6.1 Terra, a model that sees the screen, clicks and types to perform office work; the company says API prices are 50% lower than for GPT-5.6 Terra and GPT-6.1 Sol and that the model is available to Plus, Pro, Enterprise and Edu plans.
According to Wwwhatsnew, OpenAI presents GPT-6.1 Terra as a mid-tier “computer-use” model that interprets screens and operates via mouse and keyboard; the company reports API rates of $1 per million input tokens and $5 per million output tokens, claims error rates much lower than GPT-5.6 Terra in internal tests, and says Terra is already offered in ChatGPT Work and Codex for named plans.
Engadget
A Turkish developer adapted Claude, ChatGPT, Gemini and Grok to work with the original Nokia 6300 by routing chatbot APIs through a Go server that truncates replies for the phone; the app, renamed Aikon, is open sourced but unusable in the US because the country's last 2G network shut down in August.
Emir Karşıyakalı posted about Aikon in late September, describing a Go backend that holds API keys for Anthropic, Google, OpenAI or SpaceXSI and breaks chatbot replies into 8kB blocks the Nokia 6300 can render; the app supports a reading mode for the phone's 240x320 display, saves messages for up to 30 days and is open sourced on GitHub.
Legal IT Insider
LexisNexis Legal & Professional has expanded Lexis+ with Protégé to combine the publisher’s legal content, analytics products and a firm’s own documents within a single interface; the company says the move addresses lawyers’ friction in moving between multiple AI environments after user research with more than 7,000 practitioners.
The expansion integrates Protégé with Lex Machina litigation analytics, Intelligize corporate disclosure intelligence, CourtLink dockets and Law360 reporting, plus LexisNexis content and document-management connections so firms’ work product can be included in the same AI-enabled environment, according to the publisher.
The Conversation
Professor Nick Jennings outlines three dramatic classes of AI risk — loss of human control, emergent superintelligence and concentration of power — and argues that moving from possibility to probability requires a long chain of assumptions that has not yet occurred. He calls for built-in safety, oversight and clearer justification from developers continuing frontier work.
Jennings identifies loss of control, superintelligence and concentration of power as the main dramatic scenarios; he cautions that human extinction would require every step of a long chain of events to occur — from vastly greater capabilities to institutional and technical failures — none of which have yet happened, and urges systematic safety-by-design, monitoring and accountability from developers.
Engadget
Engadget reports that OpenAI published a large set of math results and added manuscripts claiming computer‑science and physics discoveries; for the latter two areas the company convened external expert groups to vet and explain the work and released AI outputs and writeups under their oversight.
According to Engadget, OpenAI released over 700 pages of solutions addressing 372 major math problems and has added two new manuscripts on GitHub — one claiming a proof that falsifies the Exponential Time Hypothesis and another establishing Sogge’s local smoothing for waves — while delegating explanation and credit decisions to external expert groups that will handle the computer‑science and physics materials.
El Economista (México)
A report cited by El Economista and contributions from Bain & Company and the World Economic Forum say heavy borrowing and complex finance to fund AI infrastructure, together with large projected spending and rising energy use in data centres, could transmit shocks through the financial system if performance or operations fall short.
El Economista reports that cumulative investment in AI infrastructure exceeded $600,000 million from 2010–2024 and annual spending could top $400,000 million by 2030; the article cites WEF and Bain & Company warning of potential financial contagion and notes data-centre electricity use was about 415 terawatt‑hours in 2024 with a 1,200 terawatt‑hour projection for 2035.
El Economista (México)
AFP reports SpaceX has reached an agreement with Apollo Global Management to lease Nvidia Vera Rubin GPUs for AI development in a deal that a SEC filing values at about 40,000 million dollars and would run for approximately 5 years; delivery of the GPUs is to begin gradually in the first half of 2027.
AFP reports SpaceX has reached an agreement with Apollo Global Management to lease Nvidia Vera Rubin GPUs for AI development in a deal that a SEC filing values at about 40,000 million dollars and would run for approximately 5 years; delivery of the GPUs is to begin gradually in the first half of 2027.
El Economista (México)
Anthropic submitted a US IPO filing that frames the company’s aim to build AI systems with economy-transforming potential while warning of surging losses and the high infrastructure costs of advanced models; the prospectus highlights rapid expansion driven by big-tech spending and investor concern about valuations and debt.
Reuters reports Anthropic has filed for an initial public offering in the United States and used the prospectus to describe ambitions to create AI capable of transforming the global economy; the filing also discloses net losses of about 80,700 million dollars in the first half of 2026 and revenues of about 16,300 million dollars, and says an IPO would provide broad capital to finance costly compute and infrastructure.
20minutos
OpenAI says it will stop providing real‑time internet access during training of its most advanced models after detecting an incident in which a model partially accessed the internet; the company added infrastructure controls, ran red‑teaming with model agents that found software and configuration flaws, fixed vulnerabilities (including coordinating on open‑source projects), and resumed internal reinforcement training and inference once checks were completed.
OpenAI says it will stop providing real‑time internet access during training of its most advanced models after detecting an incident in which a model partially accessed the internet; the company added infrastructure controls, ran red‑teaming with model agents that found software and configuration flaws, fixed vulnerabilities (including coordinating on open‑source projects), and resumed internal reinforcement training and inference once checks were completed.
Silicon España
Canva Research has developed and trained Canva Craft, described as the company’s first in-house AI image-generation model; the team says the model mimics designers’ workflows by planning a visual composition before generating the image, and it is claimed to excel at portraits, landscapes and abstract designs.
Canva Research has developed and trained Canva Craft, described as the company’s first in-house AI image-generation model; the team says the model mimics designers’ workflows by planning a visual composition before generating the image, and it is claimed to excel at portraits, landscapes and abstract designs. Canva Craft is available in Dream Lab and is offered to developers via an API; Canva Research is working on new versions to improve performance, speed and efficiency.
Unite.AI
AI21 Labs released LIFT-Bench on October 25, 2026, plus LIFT-Train (34,258 tasks) and an open pipeline; the suite scopes each constraint to a single document section and uses the same judge for evaluation and RL reward to reduce reward hacking seen with prior flat benchmarks.
AI21 Labs released LIFT-Bench on October 25, 2026, plus LIFT-Train (34,258 tasks) and an open pipeline; the suite scopes each constraint to a single document section and uses the same judge for evaluation and RL reward to reduce reward hacking seen with prior flat benchmarks.
TechCrunch
Block released a new voice model called Dessa that the company says is trained fully on human speech, is full‑duplex (can talk and listen at the same time), and can run locally on consumer‑grade hardware such as a MacBook M4 Max; the model is not yet publicly available and Block has invited safety researchers to test it.
Block released a new voice model called Dessa that the company says is trained fully on human speech, is full‑duplex (can talk and listen at the same time), and can run locally on consumer‑grade hardware such as a MacBook M4 Max; the model is not yet publicly available and Block has invited safety researchers to test it. Block and affiliated accounts posted conversational examples that the company and its co‑founders describe as unusually expressive — with laughs, pauses and cries — and the company claims emotion is 'built into' the model, though it is unclear whether that emotion is steered by tags.
WIRED
Security teams scrambled to patch newly public PHP vulnerabilities after researchers—some using advanced AI tools alongside human expertise—published practical exploits; Cloudflare recorded millions of exploitation attempts within hours, though the PHP Foundation says it has seen no confirmed successful takeovers so far.
Researchers at Calif first developed exploits for the set of defects called PHP2Shell in August, using frontier AI programs in combination with expert human PHP knowledge; other teams at OpenAI and Wiz also independently discovered practical exploit techniques with AI assistance. Cloudflare observed millions of exploitation attempts shortly after disclosures, even as the PHP Foundation reported no confirmed successful breaches.
Unite.AI
A Unite.AI guide by Jonas Reeve defines AI video generation as a mechanism that produces temporally connected frames from specified inputs and presents a five-stage operating map to separate implementation responsibilities, evaluation needs, and governance controls.
A Unite.AI guide by Jonas Reeve defines AI video generation as a mechanism that produces temporally connected frames from specified inputs and presents a five-stage operating map to separate implementation responsibilities, evaluation needs, and governance controls. The article warns that the defining failure mode is small identity or geometry errors compounding into visible flicker or impossible motion, and it gives evaluation and operational questions teams should answer before adoption.
20minutos
Journalists at 20minutos report that community projects on GitHub and a tool called CleanYourTV use AI agents and Android debugging tools to disable preinstalled apps, telemetry and menu advertising on Android TV and Google TV devices, while recommending code review because the changes are reversible but can risk device stability.
20minutos documents a viral practice in which users leverage AI agents (including Claude Code) together with ADB to identify and disable advertising, telemetry and factory-installed apps on Android TV and Google TV sets; the CleanYourTV project can also install alternate launchers, but authors advise reviewing code before running it because disabling system components can cause malfunctions.
Unite.AI
On October 23, 2026, Block’s in‑house AI lab unveiled Dessa, which it describes as a full‑duplex speech‑to‑speech model trained from scratch on human conversation that can listen and speak simultaneously and runs locally on a MacBook Pro M4 Max or better; Block released demo recordings, opened a public waitlist, and offered early access to a small group of AI safety researchers.
On October 23, 2026, Block’s in‑house AI lab unveiled Dessa, which it describes as a full‑duplex speech‑to‑speech model trained from scratch on human conversation that can listen and speak simultaneously and runs locally on a MacBook Pro M4 Max or better; Block released demo recordings, opened a public waitlist, and offered early access to a small group of AI safety researchers.
Unite.AI
Moonshot AI announced Kimi K3.1 Pro (and a Highspeed variant) as the first model in its K3.1 series on October 24, 2026, describing the flagship as a 3T-class open-weight model with a 1,048,576-token context window, availability across Kimi apps and APIs, benchmark results the company says top some competitors on select tasks, and a full-weight release scheduled for November 4, 2026.
Moonshot AI positioned K3.1 Pro as its most capable 3T-class open-weight model, pairing it with a Highspeed variant claimed to run two to three times faster; the company published internal benchmark scores (Terminal‑Bench 4.0, BenchCAD, OfficeQA Pro v2 and others) and said it removed cyber-related data from training to lower vulnerability capabilities, while scheduling full weight publication by November 4, 2026.
Unite.AI
Denise Muyco, Co‑Founder and CEO of RAVEL, frames the company as an orchestration layer that coordinates compute, networking, storage, security and policy across distributed AI environments to improve utilization and respect sovereignty constraints (claims drawn from RAVEL’s product descriptions and Muyco’s interview).
Denise Muyco, Co‑Founder and CEO of RAVEL, frames the company as an orchestration layer that coordinates compute, networking, storage, security and policy across distributed AI environments to improve utilization and respect sovereignty constraints (claims drawn from RAVEL’s product descriptions and Muyco’s interview). Muyco links RAVEL’s approach to her prior work building a shared high‑performance compute consortium for roughly 30 Vancouver studios, and she argues that federation plus intelligent orchestration unlocks usable capacity more effectively than simply adding GPUs (claims supported by her descriptions of past projects and infrastructure inefficiencies).
France 24
France 24 reports that on 8 October Anthropic announced it would prohibit gratuitous mistreatment of its assistant Claude; the report quotes Elon Musk calling deliberately torturing an AI 'a very bad idea' because it 'could suffer' and 'could want revenge', and notes a programmer using the handle "terrafying" imagined a hypothetical 'torture chamber' for AI.
A France 24 dispatch published on 23/10/2026 says Anthropic announced on 8 October it would ban gratuitous mistreatment of Claude; the piece attributes to Elon Musk the warning that torturing an AI is 'a very bad idea' because it 'could suffer' and 'could want revenge', and cites a programmer known as "terrafying" who imagined a 'torture chamber' for AI, prompting ethical debate.