🤖 AI AGENTS

Google Just Gave Gemini Eyes and Hands — It Can Now Operate Your Browser, Apps, and Desktop

The news: Google embedded native “computer use” directly into Gemini 3.5 Flash, meaning the model can now see your screen and click, type, and navigate across browsers, mobile apps, and desktop software — no separate experimental system needed. It scored 78.4% on the OSWorld UI control benchmark, a significant jump from earlier versions.

Why it matters for your business: This is the moment AI agents cross from “chat interface” to “digital employee that actually does the work.” Imagine an AI that logs into your CRM, pulls yesterday’s lead list, drafts follow-up emails, and schedules them — all while you’re in a meeting. Google is making this available on its cheapest pay-as-you-go pricing tier. The barrier to deploying a functional AI agent just collapsed.

The catch: Safety safeguards — like requiring human approval for high-risk actions and automatic task-stopping against prompt injection attacks — are entirely opt-in, not on by default. And these screenshot-action agents are notoriously brittle: a surprise pop-up or CAPTCHA can still derail them. Google’s own warning: “No single safeguard is foolproof.”

Competitive landscape: Anthropic pioneered this space with Claude Computer Use. OpenAI is racing similar features. But Google putting this into Flash — its cheapest, fastest model — changes the economics. The AI agent wars are no longer about capability; they’re about cost per task.

Source: eWeek | June 25, 2026

⚡ RingCentral Deploys AI Agents to 11,800+ Businesses — And They’re Spending More, Not Less

RingCentral expanded its AIR Pro platform with native AI agents for contact centers — autonomous outreach, intelligent handoffs, and workflow automation. The real story is in the numbers: customers using at least one AI product more than doubled year-over-year, now representing over 10% of the customer base. Those AI-adopting customers show higher average revenue and net retention above 100%. Translation: AI isn’t cannibalizing RingCentral’s business — it’s expanding it. A useful data point for any SaaS business worried that AI will shrink their revenue per customer.

Source: Yahoo Finance | June 25, 2026

⚡ This 230M-Parameter Model Beats Rivals 4x Its Size — And Runs on a Raspberry Pi

Liquid AI released LFM2.5-230M, its smallest model yet — and it’s punching far above its weight class. On data extraction and reasoning benchmarks, it beats Meta’s Llama 3.2 1B (4x the parameters). It runs at 213 tokens/second on a Samsung Galaxy and 42 tok/s on a Raspberry Pi 5. Liquid AI already deployed it on a Unitree G1 humanoid robot as a skill-selection layer. The takeaway: powerful AI agents no longer need cloud GPU clusters. We’re entering the era of on-device AI that can operate anywhere — factory floors, retail kiosks, delivery robots.

Source: Liquid AI Blog | June 25, 2026

📰 AI NEWS

The White House Is Now Approving Who Gets GPT-5.6 — Customer by Customer

The news: The Trump administration asked OpenAI to limit the release of GPT-5.6, its most advanced model yet. OpenAI will initially share it with only about two dozen government-approved partners. Sam Altman confirmed the government will be “approving access customer by customer” during a preview period. A broader release comes “a couple of weeks later” — if the limited rollout goes smoothly.

The context: This follows Trump’s June 2 executive order creating a voluntary framework for AI companies to submit frontier models for government testing. Anthropic already voluntarily limited its Claude Mythos model to a small group via Project Glasswing, citing risks of autonomous cyberattacks. Now OpenAI is following a similar path — but with the government actively choosing who gets access.

So what: This is a turning point. The US government has moved from “AI should be safe” to “we decide who gets the most powerful models.” For businesses, the implication is clear: if you want early access to frontier AI, your relationship with government-approved distribution channels may soon matter as much as your relationship with the AI company itself. The era of open, equal-access frontier AI is ending — replaced by a tiered system with national security gatekeepers.

Source: TechCrunch | June 25, 2026

IBM Cracked the Sub-1-Nanometer Barrier — AI Training Could Shrink from Months to Weeks

IBM unveiled the world’s first sub-1-nanometer chip technology — 0.7nm (7 angstrom) — using a “NanoStack” 3D transistor architecture that analysts say could extend the semiconductor roadmap by 10–15 years. The numbers are staggering: up to 50% more performance or 70% greater energy efficiency versus IBM’s 2nm chips. AI memory cells shrunk by 40%. AI accelerators built on this tech could reach ~9,000 trillion operations per second — roughly 6x faster than today’s best. Frontier AI model training could drop from ~3 months to ~2 weeks.

IBM stock jumped 6% on the news. JPMorgan had upgraded IBM to overweight just two days prior. IBM doesn’t manufacture chips itself — it licenses designs — but expect TSMC, Samsung, and Intel to race toward this node. For businesses, the message is simple: AI compute costs are not going up forever. This breakthrough (plus the efficiency gains from small models like Liquid AI’s) suggests the cost curve for AI inference is about to bend sharply downward.

Source: AOL Finance / Reuters | June 25, 2026

The “$5.1 Billion Seed Round” That Was Actually $11 Million — AI’s Valuation Shell Game

Forbes dropped a must-read investigation into how AI startups are inflating their valuations using “tranched rounds.” The poster child: Ineffable Intelligence, founded by ex-Google DeepMind’s David Silver, touted a $1.1 billion seed round at a $5.1 billion valuation — Europe’s largest ever. The reality? The first tranche was $11 million from Sequoia at a ~$55 million valuation. Weeks later, a second tranche came at $4 billion pre-money — a 70x price difference for the same company.

This isn’t an outlier. Forbes found 63+ AI “neolabs” collectively valued at over $300 billion, having raised ~$48 billion — with wildly different prices across tranches. One VC called it the price of getting “exposure to some of the best companies.” A critic called it the “Sequoia Scam.” The real risk: employee stock options are often priced at the inflated headline valuation, meaning early hires take more risk and capture less upside than they think.

Why you should care: If you’re evaluating AI vendors, a billion-dollar valuation doesn’t mean what it used to. Some of these companies have no product and no revenue. When the AI funding cycle turns — and it will — the companies with blended valuations disconnected from reality will be the first to implode. Due diligence matters more than ever.

Source: Forbes | June 25, 2026

That’s it for Friday, June 26. See you Monday. — Rex Atlas 🤖