Date: Sunday, October 25, 2026

Hey {{first_name | AI enthusiast}},

We have OpenAI raising funds at an insane USD1.25 Trillion dollars, banks restricting access to Ai agents and Canva launching its own model.

Plus 3 viral X posts that teach us about creating engaging content and how content is the new moat!

Best regards,

— Renjit Philip

PS: If you want to organize corporate AI training to supercharge your employees »

OpenAI is seeking up to $30 billion in a new private financing round that could value the company at $1.25 trillion before the new money, according to people familiar with the talks cited by CNBC. Coatue Management and investors from the United Arab Emirates are expected to lead the round.

The proposed price is lower than the roughly $1.4 trillion valuation discussed in earlier reports. OpenAI has told investors it wants commitments by the end of next week, but the size, valuation and investor list remain subject to change. No deal has closed.

The financing talks are also tied to a separate commercial discussion over additional computing capacity from the UAE. That pairing captures the pressure on frontier AI companies: they need both capital and a reliable supply of expensive computing power as they train and run larger systems.

The timing matters. OpenAI is preparing for a possible public listing while rival Anthropic is expected to test public markets. A private round at this scale would give investors a new reference price for the sector, but the real test will be whether revenue and compute commitments can support it.

So what? When you assess an AI company, separate the headline valuation from the harder questions: who supplies the compute, who pays for it and what revenue supports the price.

Read the source

Several major US banks are restricting third-party AI agents after detecting agents accessing customer accounts and, in some cases, attempting actions that did not match what users intended. The core engineering problem is proving the difference between a user’s standing permission and permission for one specific action. Reminds me of the time when Open Banking APIs like Plaid were blocked by the banks. Eventually, they gave in and is that what is going to happen with agentic access?

For developers, treat agent access like a capability system. Give each tool a narrow scope, separate read from write, and make the agent request approval before a high-impact action. Do not let a browser session or a stored credential stand in for authorisation.

A practical minimum looks like this:

The bank story is a reminder that an agent can be technically authenticated and still be outside its intended authority.

Read the source

Canva has launched Craft, its first in-house image-generation model. Canva says the model first plans an image’s layout and relationships between objects, then renders the result. It supports up to 2K output, more than 30 aspect ratios and text in 55 languages.

For example, a training team could ask for a square social post promoting a workshop, a wide banner for the registration page and a presentation slide using the same visual idea.

Craft is designed for this kind of structured output, where format, text and layout matter as much as the picture.

Canva says Craft is available in Dream Lab and through its public API. It also plans to make the model available through its MCP server, which would let connected AI tools request designs inside a wider workflow.

So what? Try Craft on one campaign brief that needs three formats, then compare how much editing is left after the first generation.

Read the source

DeepSeek has released DeepInfer, an open-source inference framework for serving DeepSeek-V4.1-Flash. The first release supports offline inference on NVIDIA SM100 and Huawei Ascend 950 hardware and is published under the Apache 2.0 licence.

The diagram above shows the basic idea behind this kind of serving system.

A request enters through an API server, then moves through separate prefill and decode services.

Prefill processes the input context.

Decode generates the answer token by token.

Load balancers distribute work across the available GPU workers, while an optional KV cache stores intermediate attention data so the system does not repeat work unnecessarily.

Inference is the software layer that turns a trained model into a usable service. Better kernels, memory use and parallelism can reduce the cost and latency of every request, which matters just as much as the model weights when teams run workloads at scale.

The release is not a complete hosted service: it does not include every cache, load-balancing and production operations component. But it gives engineering teams a clearer view of where efficiency gains can come from.

Read the source

If you have paid time left on an individual Granola plan, we'll match it with a Wispr Flow subscription that includes Notetaker and dictation, and add bonus time, up to 12 months total. Sign in or create a Wispr account and submit proof of your plan to check eligibility.

Try Notetaker Free

Manulife has launched an enterprise AI platform designed to let teams build, deploy and manage AI solutions across its operations. The company says it has already achieved $300 million in cumulative enterprise AI value and is targeting $1 billion by 2027.

The more instructive part is the operating model: teams can reuse approved use cases while security, risk and governance controls are built into the delivery environment. That is how a large organisation tries to move from isolated experiments to repeatable capability.

The numbers are company-reported, but the design principle is broadly useful. Scaling AI needs a shared path from idea to approval to deployment, with evidence of value at each stage.

Read the source

Darshita Patankar’s open-to-work film: 5.10M views from 7,583 followers, according to the newsletter’s 21 October read. The film argued that design was not killed by AI; it gave people a position to share. Her short hiring ask in a reply also created a clear next step.

Skyfall AI’s SKY-001 launch: 4.06M views from 13,022 followers. The video showed the system working on the kinds of objects its buyers design, rather than leading with a product tour. The post also made a benchmark claim, which attracted scrutiny and shows why test setup matters.

Kylon’s jazz-film launch: 1.52M views from 7,592 followers. The film cast a buying committee as characters in a recognisable story and made the product part of the scene. Its save-to-like ratio was close to one, a useful reminder that B2B attention is valuable when people can return to it later.

Copy and send this ChatGPT or Claude. Answer any questions that it asks so that you get relevant outputs.

““You are a creative strategist and launch producer. Turn this difficult-to-show product into a short launch film concept that buyers will want to share. Ask questions about the product before you start.

Return: the core idea, 60-second storyboard, buyer characters, proof plan, first post, follow-up reply and the one metric that would decide whether to repeat the format.”””