Ep. 03 – AI Assistant in Adobe Creative Cloud

Apple Price Increases and AI Infrastructure

The hosts discuss recent price rises across Apple computers and iPads. Examples include the MacBook Neo increasing from $599 to $699, the iPad Air rising by $150, and some Mac models becoming several hundred dollars more expensive. The Neo’s $100 increase is roughly 17% of its original price.

At the time of the discussion, iPhone prices had not risen. The hosts speculate that Apple might absorb some component costs to protect its market share.

They suggest that expanding AI data centers may be contributing to higher prices for RAM, GPUs, storage, and other components. Data-center operators may be competing with consumer-device makers for limited supplies and willing to pay more. The hosts also consider whether new chip suppliers and increased production could eventually ease prices, but expect pressure to continue in the near term. These explanations and forecasts are presented as analysis and speculation, not established conclusions.

Adobe’s In-App AI Assistant: First Impressions from Beta Testing

Adobe is rolling out a Firefly AI assistant in beta versions of Premiere, Photoshop, and Illustrator. Nick tested it in Premiere and Photoshop and described the experience as promising, but not yet reliable enough for general use.

Examples from the tests include:

• In Premiere, the assistant correctly labeled live-action clips and screen recordings with different colors.
• When asked to add more headroom, it adjusted the framing but moved the image too far, leaving black space in the frame.
• It identified screen-recording clips and placed them on a timeline. When asked to remove pauses, it used a transcript-based process and achieved a roughly successful, but imperfect, result.
• In Photoshop, it extracted text from a flattened graphic, but the recreated text did not match the original font and style.

The hosts see potential in automating repetitive tasks such as organizing footage, labeling clips, trimming pauses, and carrying out multi-step edits based on a stated goal. They emphasize that the beta still needs better task-specific capabilities and judgment.

They also raise a broader question: will AI help creators make their own creative decisions more efficiently, or lead them to spend more time directing AI? Their preferred outcome is for AI to assist with production tasks while creators retain control of the work.

Fable 5 and AI Safeguards

The hosts discuss Anthropic’s Fable 5 becoming available again with additional safeguards, particularly around cybersecurity. They describe classifiers that may refuse certain requests or redirect users to another model. One example is a refusal to help identify software vulnerabilities, even when the user claims to be a developer working on their own application.

The hosts consider the challenge of distinguishing legitimate defensive security work from requests that could enable exploitation. They also discuss whether governments should review model releases and whether regulators can keep pace with rapidly advancing AI.

They frame AI security as an arms race: the same capabilities can help defenders find flaws or help attackers exploit them. They argue for oversight and safeguards while acknowledging the difficulty of setting rules without slowing beneficial innovation.

MCP Connectors and Connected AI Workflows

MCP is described as a standard that lets AI assistants connect to external apps and services, such as Gmail, Google Calendar, Salesforce, HubSpot, and project-management tools. This allows an assistant to do more than answer questions: it can retrieve information or take actions. For example, it might identify a priority email, add a new client to a CRM, or answer questions using meeting transcripts and documents.

The hosts describe MCP as a way to bring information from several services into a central AI assistant such as Claude or ChatGPT. They compare its adoption to USB-C: a common interface that many companies and tools can support.

They note that Claude and ChatGPT offer connectors, while Microsoft Copilot’s support was described as less standardized at the time. Microsoft’s existing integration with Microsoft 365 may reduce its immediate need for external connectors.

The hosts also discuss practical considerations:

• Connector availability varies. They cite roughly 200 connectors in Claude and more than 8,000 integrations through Zapier.
• Access can often be limited by permission, such as allowing an assistant to read email without giving it permission to send messages.
• Claude can require confirmation for actions or let users set rules for how actions are approved.
• Connecting more services increases an assistant’s usefulness, but also makes security and careful permission management more important.

The hosts envision AI assistants becoming a primary interface for tasks: users state a goal, and the assistant uses connected services to complete it. They caution that users and organizations should consider what data and permissions they grant.

Reviews, Guides, and Tutorials Mentioned

• Adobe beta first impressions: Nick’s hands-on testing of the Premiere and Photoshop assistants, including examples of successful and flawed results.
• MCP demonstrations: The hosts describe finding and connecting apps through Claude’s connector settings and discuss using connectors with ChatGPT.
• Gamma: Nick says he made a video about Gamma, a presentation-creation tool, including how to use MCP to draw information from email and documents.
• Copilot Co-work: Nick says a video comparing it with Claude Co-work, including billing details, is planned for Kevin’s channel.
• ComfyUI: Kevin mentions an upcoming tutorial on generating AI videos locally with the free, open-source tool, subject to having suitable computer hardware.
• Google Flow: Kevin describes an upcoming look at Google’s AI video workflow, which develops a concept into text and visual storyboards, generates clips, and provides AI-assisted timeline editing.
• The hosts also mention forthcoming videos on Claude Code and ChatGPT.

Main Speakers and Sources

• Main speakers: Kevin Stratvert and co-host Nick Rossi.
• Sources mentioned or referenced: Anthropic, OpenAI, Apple, Microsoft, Adobe, Google, Zapier, TechCrunch, and the United Nations. The hosts also refer to their own product testing and commentary.