29 October 2024

Claude's HUGE Double Update: Computer Control + Sonnet 3.5 (New)

2Frameworks

9Insights

Frameworks in this episode

Productivity6 steps

Outcome-Verifiability Model Selection

Match AI models to how clearly their outputs can be judged

Read →

Strategy6 steps

Trajectory and Second-Order Implications Lens

Judge an early product by its direction and downstream consequences

Read →

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Hot Take· 3

Hot Take03:00

AI Assistants Could Push Business Software Into the Background

The hosts argue that computer-using assistants may become the primary interface to business software. Instead of learning and administering many tools, workers could express an objective and let an assistant navigate the underlying applications.

The launch's primary demonstration is B2B rather than consumer-focused
AI may sit above software as the user-facing interaction layer
Workers could spend less time administering tools
Software may increasingly be designed for machine users

“software being un moved into the background because the knowledge assistant is able to do all of the use cases do the thing you want…”
Kieran · 03:30

“so what happens when the core buyer and user of the B product is the knowledge assistant itself”
Kieran · 04:30

#b2b#software#knowledge-work#ai-agents

Hot Take04:30

From Word of Mouth to Word of Machines

As assistants gain purchasing authority, machines may evaluate, buy, install, and operate products for other machines. That shift could change how vendors build reputation, communicate value, and win software purchases.

AI assistants may become both software buyers and users
Machine-to-machine decisions could influence purchasing
Traditional product marketing may need machine-readable signals
Corporate cards could eventually enable autonomous procurement

“we're moving from Word of Mouth to word of machines where machines are going to make decisions with other machines”
Kipp · 04:30

“the knowledge assistant having access to a corporate card that you give it and buying the software installing the software and using the software”
Kieran · 04:30

#procurement#machine-buyers#b2b-marketing#agents

Hot Take17:30

Built-In Platform Styles Could Make AI Content Look the Same

Claude's updated writing behavior appears to reproduce platform-appropriate styles without a manually supplied style guide. Kieran sees this as useful in the short term but warns that widely distributed defaults can rapidly saturate a style and reduce differentiation.

Claude appears better at platform-specific creative writing
Some previously manual style guidance may now be built into the model
Easy access causes many creators to adopt similar patterns
Distinctive creators may need to change their style sooner

“it started doing some styles that I had been using using style guides for without me even uploading the style guide”
Kieran · 17:30

“I should probably start switching up my style because when you've given that to everyone that style will become pretty uh saturated pretty quickly”
Kieran · 18:00

#creative-writing#content-style#differentiation#marketing

Explainer· 3

Explainer06:30

Why Anthropic Looks More Enterprise-Focused Than OpenAI

The hosts connect the companies' product strategies to their revenue mixes. They characterize OpenAI as currently more consumer-oriented and Anthropic as an infrastructure company whose API revenue creates stronger incentives to build enterprise capabilities.

OpenAI receives substantial revenue from its consumer product
Anthropic reportedly earns more through API infrastructure
Computer use aligns with enterprise workflow automation
Revenue composition can help explain product priorities

“when you look at open ai's Revenue mix they are a consumer company”
Kipp · 06:30

“entropic are an infrastructure company most of their money is coming through the API”
Kipp · 06:30

#anthropic#openai#enterprise-ai#business-models

Explainer12:00

Future Assistants May Discover and Automate Your Repeated Work

The hosts envision hardware-integrated assistants observing a person's recurring computer activity. The assistant could identify repetitive tasks, propose an agent, and build the automation after the user approves it.

Hardware access could reveal recurring workflows
Assistants may proactively suggest automation opportunities
Users could approve agent creation with minimal setup
Personal context would help determine which work to offload

“it should be able to do then is to show you on a day-to-day basis what are your most common tasks that you complete”
Kipp · 12:30

“it would say hey do you want me to build an agent to do this task and you should be able to click a button…”
Kipp · 12:30

#personal-assistant#workflow-discovery#automation#hardware

Explainer13:30

Computer Use Is Constrained by Data Access and 15-Minute Runs

The initial computer-use release has significant practical limits. It cannot yet be trained on internal data as described in the episode, depends on information visible on screen or available on the web, and can act for roughly 15 minutes before model context limitations become restrictive.

The agent cannot yet be trained on internal data
It depends on web-accessible or on-screen information
Action runs are limited to about 15 minutes
The model retains its existing context-window constraints

“you can't train computer use on internal data yet”
Kipp · 13:30

“it can only run for about 15 minutes”
Kipp · 14:00

#limitations#context-window#computer-use#internal-data

Story· 1

Story01:30

Claude Moves Vendor Data Across a Spreadsheet, CRM, and Form

Anthropic demonstrates computer use through a fictional vendor-onboarding task. Claude inspects a spreadsheet, discovers the vendor is missing, searches a CRM, and transfers the recovered information into a form by operating the interface.

Claude takes screenshots to understand the visible interface
The agent switches applications when the spreadsheet lacks the vendor
It searches the CRM for additional information
It autonomously transfers data into the request form

“what we're going to do is ask look at the spreadsheet check if an equipment is in there and if not move over to the…”
Anthropic researcher · 02:00

“then autonomously starts transferring the information across without me having to do anything”
Anthropic researcher · 06:00

#computer-use#automation#crm#vendor-onboarding

Tool· 1

Tool11:00

Replit Turns Computer Use Into an Accessible Agent Template

A Replit demonstration uses Anthropic's computer-use API to find a YouTube video and skip advertisements. The hosts emphasize that templates let non-developers fork simple examples and extend them into richer workflows involving transcripts, content drafts, or other Claude-powered tasks.

Replit provides templates that users can fork
The demo navigates YouTube and skips ads
Non-developers can experiment with computer-operating agents
Simple browsing tasks can become parts of larger content workflows

“you could do this now without having to be a developer which is pretty amazing right”
Kipp · 11:30

“really what Claude use API is is a agent builder for your laptop or your Hardware”
Kieran · 12:00

#replit#computer-use#no-code#agent-builder

Takeaway· 1

Takeaway08:00

Agents Could Turn Company Rules Into Standardized Execution

Companies may adopt agents quickly where human training and policy compliance are inconsistent. An agent can apply spending limits, CRM requirements, and other operating rules in a repeatable way, although the episode does not explore governance risks in depth.

Training and compliance are persistent operational challenges
Agents can apply the same rule every time
Expense management is an early example
Standardization may be a stronger enterprise driver than convenience

“you know what the biggest challenge is training and compliance right”
Kipp · 08:00

“I can like have a much more standard way of doing something which is actually what I want”
Kipp · 08:30

#compliance#training#standardization#enterprise-agents