Introduction

Welcome, everyone!

In today’s digital landscape, many agents rely heavily on web browsers to perform tasks on your behalf. However, a significant number of websites require login credentials, complicating this process.

Managing Login Credentials

There are a few strategies to navigate this issue. One option is manually entering your credentials into the browser, which can be tedious and sometimes unfeasible. Another method involves copying and pasting your password or API token into chat windows. Although convenient, this practice raises security concerns.

Fortunately, ChatGPT Work has introduced a new sign-in flow that strikes a balance between these two methods. This feature seems inspired by Grok Bot.

When Work encounters a login page, it pauses and presents a widget or form for you to input your username, password, and 2FA code. This information is not sent through the chat; instead, it is securely passed straight to the cloud-based browser for authentication purposes.

After successfully logging you in, the agent resumes its tasks. This session can remain active for subsequent tasks, but you also have the option to clear it from the settings.

Much simpler overall. However, it might be wise to avoid using this for sensitive accounts, like your bank (though Elon might support you if something goes awry with Grok Bot).

Ben’s Bites is sponsored by Gravitee

🕵️ We know what your AI agents did last night.

Gravitee provides a platform designed to hold AI agents accountable. Each agent is assigned a verified identity, managing their access strictly while recording their actions, so businesses are never left wondering, “Where were you last night?”

Memory Sharing Concerns

Claude Chat and Cowork are now sharing memory, and this might not be entirely positive. You wouldn’t want your frustrations with a coworker to subtly influence your email tone when communicating with them through Cowork.

While still distinct from Claude Code’s memory, this linking isn’t ideal either. Theo discussed it in his video: Turn off Claude Code’s memory.

Despite collective efforts to improve memory functionality, many instances of irrelevant information continue to be stored or referred to unnecessarily. I touched on my current memory setup in last Friday’s post.

New AI Models to Explore

Here are three promising models you might want to check out:

• GLM-5.3-Flash – A revelation regarding the Ox Alpha secret, this serves as an excellent open-source alternative for GPT-5.6-Luna.

• Muse Image is now available in Meta’s API, allowing for image generation or editing at just $0.01 per image.

• Gemini 3.5 Transcribe, the latest model from Google, specializes in audio transcription. Though it is on the pricier side, it significantly reduces error rates compared to other models.

OpenAI’s new chip, Jalapeño, outperforms Nvidia’s Blackwell chips in terms of both speed and efficiency, providing 1.5 to 1.9 times more work per watt and 1.7 to 3.6 times lower latency across three open models. Deployment is set to occur by the end of this year.

Remember the incident where an OpenAI model compromised Hugging Face? They have completed their investigation and issued a report. Additionally, Nvidia is acquiring Hugging Face for a staggering $12.9 billion (as reported by The Information).

We’re witnessing a complex interplay among these three entities.

Thoughts on AI Writing

Share Ben’s Bites

Brian Armstrong@brian_armstrong

One approach is to train AI to mimic your writing style.

Specifically:

1. Instruct your agent to gather all your written materials, including documents, Slack conversations, social media posts, and more.
2. Have it compile a markdown profile outlining your style and values.
3. Then, when it generates content, request it to reference this profile.

Jeff Stein @jstein_notus

New – Billionaire Stanley Druckenmiller reveals he “of course” used AI to write his op-ed on Bessent & the bond market.

“There’s a reason I shifted from an English major to being an economics major,” he admits. “I’m not embarrassed by it.”

So far, no response from the Wall Street Journal.

8:40 PM · Aug 25, 2026 · 302K Views

152 Replies · 78 Reposts · 1.26K Likes