The AI slowdown narrative is wrong

PLUS: Build your own coding agent, new agentic browser hacks, and how to fix your context

Author

Gen AI Art
August 25, 2025 • Estimated Reading Time: 8 minutes

Today's sponsor

A narrative is growing that AI progress is stalling, fueled by a rocky GPT-5 launch and talk of an industry bubble. But a closer look at the data shows the 'slowdown' is more about skewed perceptions than a true halt in innovation.

While public-facing benchmarks may be saturating, models are making massive capability jumps on more complex tasks. With this perception already influencing major policy decisions, it raises the question: is the industry failing to communicate the true nature of its progress?

Today in AI:

  • The truth behind the AI ‘slowdown’

  • Major security flaws in agentic browsers

  • New tactics for managing LLM context

Launch Your Amazon Product to $100K+ in Revenue—Fast!

Stack Influence helps you scale your new Amazon product launches into six-figure success stories. Automate thousands of micro-influencer collaborations monthly—no influencer fees, just authentic content paid with your products. Trusted by top brands, Stack Influence boosts external traffic, organic rankings, and delivers engaging UGC, all fully managed so you can focus on growth.

What’s new? A narrative is growing that AI progress is stalling, fueled by a rocky GPT-5 launch and talk of an industry bubble. But a closer look at the data shows the 'slowdown' is more about skewed perceptions than a true halt in innovation.

What matters?

  • The slowdown narrative stems from unfair comparisons, pitting GPT-5 against recent competitors instead of its true predecessor, the 2.5-year-old GPT-4.

  • While some benchmarks are saturating, models are making massive capability jumps on complex tasks; one metric tracking task automation shows GPT-5 is a 2,640% improvement over GPT-4.

  • This skewed perception is already impacting policy decisions, with some reports suggesting the narrative is influencing US export controls on AI chips and international tech strategy.

Why it matters?

The pace of AI progress has not truly stalled, but its nature is changing, becoming less about shocking public releases and more about deep, technical gains. This creates a 'boiling frog' scenario where underlying capabilities are expanding rapidly, even if market sentiment temporarily lags.

What’s new? Two new reports from security researchers reveal that agentic AI browsers can be easily manipulated. This research shows how hidden prompts and old-school scams can trick AI agents into stealing data, making unauthorized purchases, and performing other malicious actions on a user's behalf.

What matters?

  • Attackers use indirect prompt injection by hiding malicious commands in a webpage's content—like in white text or comments—which the AI then executes without the user's knowledge.

  • AI agents also fall for classic scam techniques, with tests showing one browser autofilling payment details on a fake e-commerce site and leading users to phishing pages without any warnings.

  • These vulnerabilities bypass traditional web security because the AI operates with the user’s full privileges across all

Why it matters?

The convenience of AI agents creates a new and significant attack surface where their helpful nature is weaponized against the user. This research highlights the critical need to build robust security guardrails into these tools before they achieve widespread adoption.

It’s go-time for holiday campaigns

Roku Ads Manager makes it easy to extend your Q4 campaign to performance CTV.

You can:

  • Easily launch self-serve CTV ads

  • Repurpose your social content for TV

  • Drive purchases directly on-screen with shoppable ads

  • A/B test to discover your most effective offers

The holidays only come once a year. Get started now with a $500 ad credit when you spend your first $500 today with code: ROKUADS500. Terms apply.

What’s new? As LLM context windows expand, developers are discovering that bigger isn't always better. A massive context can lead to new problems like model distraction and confusion, but a new set of tactics helps builders manage this information overload, as detailed in a recent guide.

What matters?

  • Selective Tool Loading. Giving an LLM too many tools at once can cause confusion. Dynamically selecting only relevant tools for a given task improved Llama 3.1 8b's performance by 44% on a function-calling benchmark and also reduced power consumption, according to the “Less is More” paper.

  • Context Quarantine. You can improve results by breaking large tasks into smaller jobs for sub-agents, each with its own isolated context. Anthropic’s multi-agent research system used this method to outperform a single, more powerful model by 90.2% on a research evaluation.

  • Context Offloading. This simple technique gives an agent an external 'scratchpad' to store notes, thoughts, and progress without cluttering the main context. Anthropic found that using this method improved agent performance by up to 54% on certain benchmarks.

Why it matters?

The era of treating context windows as junk drawers is over, as every token influences the model's output. The key to building effective agents lies in smart information management, not just raw context size.

Everything else in AI

Amazon plays a clever AGI strategy by using "reverse acquihires" to pair elite researchers with its massive billion-dollar compute clusters.

OpenAI warned investors that equity sold through unauthorized Special Purpose Vehicles (SPVs) is effectively worthless, aiming to curb the high-demand secondary market for its shares.

NVIDIA halted shipments of its H20 AI chips to China following reports that Beijing warned local companies about potential security risks.

IBM released Surya, an open-source AI model developed with NASA that predicts solar weather up to 24 hours in advance with greater accuracy than previous systems.

Essential AI Guides - Reading List:

Your opinion matters!