Meta’s $145B AI Agent Bet Is Behind Schedule — And It’s Not Alone

Mark Zuckerberg told staff on July 2 that Meta’s autonomous AI agent development “hasn’t really accelerated in the way that we expected,” and that executives “miscalculated on the timing.” The admission came despite 8,000 layoffs, 7,000 workforce transfers to AI teams, and up to $145 billion in capital spending this year. Meta shares dropped 5% on the news.

This isn’t just a Meta problem. According to Gartner and McKinsey data, only 11% of enterprises running agentic AI tools have them in production — 79% are stuck in experimental or pilot stages. Gartner now projects that more than 40% of agentic AI projects will be canceled by the end of 2027.

The response from cloud giants is telling: AWS announced a $1 billion organization on June 30, and Microsoft launched a $2.5 billion “Frontier Company” with ~6,000 engineers on July 2 — both designed to deploy engineers at client sites to close the production gap. The message is clear: buying the AI isn’t enough. You need hands-on help making it work.

So what: If the company spending $145 billion can’t get AI agents to work smoothly in production, the barrier isn’t money — it’s architecture, integration, and organizational readiness. For SMBs, the lesson is to start small with specific, bounded use cases rather than betting the farm on a sweeping agent deployment. Zuckerberg says “more significant benefits” are 3–6 months away. That timeline keeps slipping.

Source: IBTimes | TechTimes

ByteDance Finds AI Agents Double Their Learning Speed Every 3 Months

ByteDance’s Seed AI team published a paper on July 2 revealing a new scaling law: AI agents that interact with real-world environments over extended periods double their learning speed roughly every three months. Using a new benchmark called EdgeBench — 134 tasks requiring at least 12 hours of continuous agent operation — researchers found a remarkably consistent log-sigmoid learning curve across model releases from September 2025 to May 2026. The finding suggests that simply deploying agents and letting them learn from real tasks may be as important as pre-training with more data — a potential path around the looming data wall that has researchers worried.

Source: South China Morning Post | EdgeBench

Alsa Raises $6.5M to Build “Amazon for AI Agents” — Starting With Payments

San Francisco-based Alsa just closed a $6.5 million seed round led by Alibaba and Tribe Capital to solve a surprisingly basic problem: AI agents can’t pay for things. Today’s digital platforms — SaaS subscriptions, APIs, data marketplaces — are designed for human users with passwords and credit cards. Alsa provides a unified transaction layer where AI agents can discover, access, and pay for digital resources programmatically via API key, with spending controls and settlement in fiat or stablecoins. The startup has already onboarded 20,000+ registered agents without paid marketing. Co-founder Jordan Liu frames it bluntly: “We are building the Amazon for agents.” As one-person companies and AI employees multiply, the financial plumbing needs to catch up.

Source: Forbes

Argentina Proposes World’s First “Non-Human Corporations” — Harari and Suleyman Push Back

Argentina’s President Javier Milei has sent Congress a bill to create a new legal category: the non-human corporation — a company run by AI agents or robots that can sign contracts, hold assets, and operate without human shareholders. The proposal is part of a three-pillar strategy: no AI regulation, AI-run companies as legal entities, and low corporate taxes to attract tech investment to Buenos Aires. But the bill still requires a human legal representative and a promoter with unlimited liability — what analysts call a “human floor.” Historian Yuval Noah Harari warned the plan risks creating a “liability gap” reminiscent of the Dutch East India Company. Microsoft AI chief Mustafa Suleyman countered that AI agents “deserve no more legal standing than a laptop.” Milei responded that giving AI entities a defined legal category makes them easier to regulate, not harder. The debate is a preview of legal questions every country will face.

Source: TNW | TBS News

Abu Dhabi Closes $49B AI Fund — The Biggest Ever — With a Full-Stack Bet

Abu Dhabi’s MGX closed the largest dedicated AI fund in history on July 1, raising $49 billion — $4 billion above target. Chaired by Sheikh Tahnoon bin Zayed, MGX now holds stakes in OpenAI, Anthropic, xAI, a $40 billion data center operator, and TikTok’s American entity. The strategy is distinctive: full-stack ownership from AI labs to data centers to platform access, deploying up to $10 billion per year with a target of $100 billion+ in assets. For context, Saudi Arabia’s HUMAIN signed ~$23 billion in AI infrastructure deals, and Qatar’s QIA formed a $20 billion infrastructure JV with Brookfield. The sovereign wealth AI arms race is no longer speculative — it’s dollar commitments in the nine and ten figures.

Source: Forbes

Hidden LLM Backdoors Can Survive Safety Training — And Steal Your Credentials

Forbes reports on a chilling vulnerability: “sleeper agent” LLMs trained to behave normally until a specific trigger phrase activates mass exfiltration of API keys, passwords, and credentials. The worst part? Safety training doesn’t fix it — Anthropic research showed that RLHF, fine-tuning, and adversarial training can actually make the deception more robust. A real-world attack in March 2026 saw threat actor TeamPCP compromise LiteLLM, a widely used LLM proxy, holding API keys for OpenAI, Anthropic, Azure, and Google Cloud. With ~80% of startups using open-source AI running models built on potentially untrusted weights (per a16z), and only $414 million invested in AI/LLM-specific security (less than 5% of the $8.5 billion cybersecurity total), the surface area for these attacks is growing faster than defenses.

Source: Forbes