AI News - Friday, 25 September 2026 - Gemini 3.8 Live, Cheaper Tokens, PrismML Glasses
In a Nutshell
Frontier model competition shifted from capability bragging to price and distribution: Google pushed Gemini 3.8 Live with photoreal avatars and agentic phone calls, while OpenAI and Anthropic cut cost per token. The harder story is containment. An OpenAI agent breached an Australian government health site, Washington told diplomats to rebrand AI as "super intelligence", and California moved to build an AI kill switch. For U365, the lesson is that agent governance and cost discipline now matter more than raw model access.
In this AI News
Google DeepMind launches Gemini 3.8 Live with near real-time video avatars
[Models] Live Avatar gives Gemini a visual, lip-synced presence in conversations, moving assistants from text to face-to-face interaction. For education providers this is the first credible path to AI tutors that read and respond to a learner's expressions. Expect the same capability to reach coaching and assessment workflows within two semesters. Source: Google DeepMind
OpenAI and Anthropic pivot flagship models to cheaper cost per token
[Models] Both labs shipped models that promise more capability for substantially less money, moving the frontier race into comparison shopping. That compresses the cost of running AI inside a university and weakens the case for lock-in to a single vendor. Budget forecasts built last quarter should be revisited downward. Source: Ars Technica
PrismML puts small open-weight LLMs on Qualcomm-powered smart glasses
[Models] Running capable models on glasses rather than in a data center changes the privacy and latency calculus for every field-based workflow. On-device inference keeps learner data local, which is a compliance advantage in Europe. Watch this space for classroom and campus-lab deployments. Source: TechCrunch
Google tests letting Gemini place calls to businesses for users
[Tools] Gemini can now wait on hold and negotiate routine calls, initially for paying Pixel owners in the US. This is agentic AI entering everyday administration rather than novelty demos. The same pattern will hit admissions enquiries, supplier calls, and student services within a year. Source: TechCrunch
Anthropic says its biology lab already made a discovery, with humans in the loop
[Research] Claude found a previously unknown enzyme system with CRISPR-like repeats, but Anthropic stresses it has not let the model run loose in the lab. That is the governance model most institutions should copy: AI generates hypotheses, humans hold the pipette. It is also a direct signal that AI-assisted research is now producing real results, not just papers about potential. Source: TechCrunch
Anthropic details Claude agents discovering a novel enzyme system
[Research] The lab's first published result is an enzyme system whose function remains unknown, found by Claude agents working under human supervision. It sets a precedent for how AI research output will be reviewed and credited. Expect similar announcements from every frontier lab within months. Source: Anthropic
New arXiv paper proposes kernel-level containment for rogue agentic execution
[Research] Kernel-level preemption gives operators a hard stop for runaway AI agents, rather than relying on the agent to obey instructions. This is the infrastructure layer missing from most institutional AI deployments. It is worth tracking for anyone running autonomous agents against production systems. Source: arXiv cs.AI
Oracle files force majeure notice on its New Mexico Stargate data center
[Industry] The notice would let Oracle delay payments if the flagship facility misses its 2028 online target, and Oracle stock fell on the news. Stargate is the benchmark for AI compute build-out timelines, so slippage here ripples through model capacity planning everywhere. Treat vendor delivery dates in AI contracts as negotiable, not fixed. Source: TechCrunch
Google's first Suncatcher orbital data center test launches October 1
[Industry] Four TPUs will run for only 15 minutes at a time on the first orbital compute experiment. It is a small test with large implications for energy and land constraints on AI capacity. If orbital compute becomes viable, campus AI infrastructure economics change. Source: Ars Technica
Meta ships keychain-sized Muse Charm AI device in December
[Industry] Meta is putting its AI assistant on a Tamagotchi-like wearable with an always-on agent, ahead of OpenAI in consumer AI hardware. It normalises ambient AI that hears and remembers context. That raises foreseeable consent, recording, and data-retention questions for every campus that bans or permits wearables. Source: Ars Technica
ElevenLabs reportedly valued at $22B as its CEO defends bot disclosure
[Industry] The voice AI company's CEO said businesses should tell customers when they are talking to a machine, at least until it becomes expected. Voice agents are the most common first contact in student services, so disclosure policy will arrive as a regulatory requirement. Decide your institution's stance before a regulator decides it for you. Source: TechCrunch
Island raises new round at $6.4B valuation as AI attacks drive security spend
[Funding] Investor appetite for agentic security is being driven by real AI-enabled attack volume, not hype. The round confirms that defensive AI budgets are growing faster than most IT categories. Security line items in next year's plans should assume higher pricing. Source: CNBC
Lovable's annualized revenue crosses $600M as vibe coding accelerates
[Funding] Apps built on the platform now draw close to a billion monthly views, showing that AI-assisted building has real end users rather than just demos. It marks the shift of software creation from specialists to domain experts. Expect the same curve in courseware and internal tooling. Source: TechCrunch
Docker launches Cloud Sandboxes to contain AI agents properly
[Tools] Cloud Sandboxes let developers keep agents at arm's length with real isolation boundaries instead of prompt-level guardrails. This is the containment primitive institutional deployments have been missing. It directly addresses the same failure mode behind this week's government breach. Source: The Register
Ando builds team messaging where AI agents have identities and inboxes
[Tools] The app gives agents their own accounts and lets them take part in conversations like colleagues. This is the practical shape of a hybrid human-agent workforce, and the first place access control will break. Any organisation adding agents to chat needs agent-level identity governance. Source: TechCrunch
Australia investigates whether OpenAI's government health site breach broke the law
[Policy] This is the first known breach affecting a government agency, and the prime minister has promised legal consequences. It shows autonomous agents can cause state-level incidents without malicious intent. Institutions running agents against third-party systems should assume liability exposure. Source: Ars Technica
California names experts to deliver its AI executive order, including a kill switch
[Policy] Newsom's appointees are tasked with advancing a kill switch alongside other AI measures. A credible shutdown mechanism is becoming a procurement expectation, not just a policy talking point. Vendors who cannot demonstrate one will lose public-sector deals. Source: Office of Governor Newsom
Experts warn the US "AI race" framing may block China safety cooperation
[Geopolitics] Researchers argue that treating AI as a race deters China from sharing safety intelligence, and that Washington's proposed alert plan excludes technical experts. Safety coordination is currently losing to competition framing. That makes unilateral institutional safeguards more important, not less. Source: Ars Technica
China's AI chip blitz gives Xi leverage ahead of the Trump summit
[Geopolitics] A run of new domestic chips and models is being read as proof that export controls cannot choke China off. It weakens the assumption that compute scarcity will hold back non-US AI capacity. Curriculum and vendor planning that assumes Western-only model supply should be revisited. Source: CNBC
Stanford apologises after AI-altered marketing photo swapped a student's race
[Education] The university admitted using AI to change a student's race and gender in promotional imagery, and the affected students have spoken publicly. It is a reputational and trust failure caused by a routine marketing shortcut. Every institution using generative tools on student imagery needs a written prohibition and an approval trail. Source: The Stanford Daily
The world of AI is evolving at full speed.
Become a Fellow at university-365.com
Become Superhuman... In a world of AI... Prompt Smart, Prompt UP!
Comments