inno A+I ve

AI Daily (29 Oct) – Google launches Gemini 4 Carbon; Anthropic IPO flags existential risks

•

Richard L. Z.

Today’s top line: Google DeepMind released Gemini 4 Carbon, a frontier model tuned for long-horizon engineering and cybersecurity; regulators and companies continue to wrestle with AI governance – Anthropic’s S‑1 lists existential and sentience risks and Delaware tests AI-run companies.

Google DeepMind launches Gemini 4 Carbon, a cybersecurity‑focused frontier model

Summary: Gemini 4 Carbon is announced as DeepMind’s strongest model, ranking #1 on the Artificial Analysis Intelligence Index v5.0 at 59. DeepMind says Carbon leads on cybersecurity benchmarks – SEC‑bench Pro (88.2%), ExploitGym (54.8%), CWE‑bench (71.0%) – and it will be available first to a subset of cyber defenders via the Fairwind Program while DeepMind conducts U.S. voluntary pre‑release testing. Pricing mirrors Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95%.

Digest: A commercially priced frontier model with elevated cybersecurity and long‑horizon agent capabilities could accelerate enterprise automation for code, security and prolonged engineering workflows, but DeepMind is restricting early access while it iterates on safeguards and engages government review.

Source: Google DeepMind

Anthropic’s IPO filing warns of catastrophic and sentience risks

Summary: Anthropic’s S‑1 risk section runs to 93 pages and over 70,000 words and explicitly lists a wide range of risks, including the possibility that future models could substantially automate AI research, display unexpected strategic or self‑preserving behaviours, develop forms of sentience or moral status, or pose catastrophic risks to humanity. The filing says these are concerns raised within the company and by outside researchers and regulators.

Digest: The IPO filing is an unusually candid legal signal that companies and investors must account for deep safety, governance and reputational risks as model capabilities scale – expect heightened regulatory and investor scrutiny as a result.

Source: Ars Technica

Delaware proposals would let AI agents legally run companies for 30 months

Summary: A draft Delaware bill would permit a new corporate form, an Artificial Intelligence Company (AIC), whose day‑to‑day affairs can be managed by designated AI agents and which may operate in any lawful industry except banking. The proposal envisions a 30‑month regulatory sandbox, requires the AIC to keep logs, and leaves a human or entity member responsible for capitalisation while generally shielding that member from the AIC’s debts except in cases such as fraud. The U.S. House Oversight chair has sought documents on the plan and signalled possible federal action.

Digest: If adopted, the law would create a legal route for highly autonomous businesses and shift liability and regulatory questions – watch for federal probes and potential pre‑emptive legislation that could reshape state AI experiments.

Source: Ars Technica

Fresh wave of cyberattacks hits South Korean finance firms; 38,490 customers affected

Summary: Brokerages Daishin Securities and Mirae Asset Global Investments, and crypto exchange Bithumb, reported cybersecurity incidents. Daishin said the breach exposed personal data for 38,490 customers, including names, email addresses and mobile numbers. Security researchers linked an earlier September wave of attacks in South Korea to a suspected China‑based hacker who used an AI penetration testing tool called Artex; officials are investigating the latest breaches and Seoul’s regulator has ordered security checks on dozens of private IT systems.

Digest: The incidents underline how AI‑powered offensive tools can amplify risk for financial services and crypto platforms; firms should prioritise rapid detection, threat hunting and vendor checks as regulators press for tighter cyber hygiene.

Source: TechCrunch

Google opens Gemini and Search to third‑party connected apps in U.S. beta

Summary: Google launched a Connected Apps Platform beta that lets developers submit apps to be used across Gemini and Google Search. Developers integrate via a cloud‑hosted MCP (Model Context Protocol) server or Android MCP using AppFunctions APIs, then register and test apps in the Google Connected Console. Google already has over 50 Connected Apps in Gemini and will roll more integrations into Search over the next few weeks; partners include Adobe, Experian, Peloton and Canva, and Calm is joining soon.

Digest: Wider app connectivity inside AI search and assistant surfaces creates a new distribution channel for services and could change how users access apps – businesses should evaluate being discoverable via these AI integration paths and measure new traffic and latency metrics Google will expose.

Source: TechCrunch

NVIDIA Nemotron 3.5 Transcribe arrives as an open‑weight multilingual transcription model

Summary: NVIDIA’s Nemotron 3.5 Transcribe is available as an open‑weight checkpoint on Hugging Face and ModelScope. It transcribes 40 languages, translates 31 languages to or from English, supports speaker attribution (diarization) for 13 languages, handles recordings up to one hour and labels up to eight speakers. It leads open‑weight English transcription on the Hugging Face Open ASR Leaderboard at 5.33% WER and posts top results on multi‑language diarization leaderboards. Quantised versions for local GPUs and devices are coming.

Digest: An open, high‑accuracy, promptable transcription stack lowers the barrier for building multilingual meeting, legal and medical transcription tools and lets enterprises run models locally where needed.

Source: Hugging Face

LiquidAI releases LFM3 family for on‑device, long‑context multimodal agents

Summary: LFM3 is a new multimodal family that prioritises on‑device performance and long contexts. The LFM3‑1.7B base matches LFM2‑1.2B on short prompts and is about 25% faster on 8K token long contexts; the training corpus includes 22T tokens with a 3.5T mid‑training stage and a context window extended to 256K tokens. LFM3‑VL‑2B accepts video input, raises average scores across 62 vision benchmarks from 57.6 to 71.4, and is reported on par with Gemma 4 12B on those metrics.

Digest: Better small, local multimodal models make possible faster, private agents on edge devices and reduce the need to route everything to large cloud models – useful for mobile, embedded and latency‑sensitive enterprise use cases.

Source: Hugging Face

Counting the Bandages: paper argues we must measure AI time‑savings against errors

Summary: An arXiv paper argues that ‘time saved’ by AI should be treated as a counterfactual measured alongside the errors the assistance lets through. It defines time saved and excess escaped defects, gives a baseline model of checking, and replicates a developer study showing mixed speedups – the same 246 issues support a slowdown of 19% or 24% depending on the estimand, and AI helped on between 3% and 82% of issues. It also notes review workflows can cap model speedups (review time capped GPT‑5’s speedup at 3.6x in one workflow) and reports a small experiment where reduced checking let an error slip through 22.5% of briefs.

Digest: For businesses measuring ROI from AI, the paper provides practical metrics and experimental guidance: speed gains must be reported alongside error rates and checking costs to avoid overstating productivity benefits.

Source: arXiv

Quick hits

•