# What's new from Anthropic and what's next

Alex Albert, Anthropic | AI Engineer World's Fair 2024 | 13:38

Source: https://www.youtube.com/watch?v=EuC1GWhQdKE
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/whats-new-from-anthropic-and-whats-next
Published: 2024-08-05
Tags: design, multimodal, product-strategy, tool-use

## TL;DR
- Claude 3.5 Sonnet combines stronger coding, reasoning, retrieval, and vision capabilities with lower cost and wider platform availability.
- Artifacts and Projects show how products can be designed around language models instead of adding a chat box to an existing interface.
- Anthropic expects models to become smarter, cheaper, and faster within months, so developers should plan products that can use newer models during development.

## Summary
Alex Albert argues that AI products are still in the equivalent of replacing a steam engine with an electric motor. Teams often add a chat button to an existing product instead of redesigning the workflow around what language models can do. He uses Claude 3.5 Sonnet, Artifacts, and Projects as examples of a different approach. Sonnet improves coding, retrieval, vision, and multi-step work, while costing less than Claude 3 Opus and being available through several cloud platforms. Artifacts separates generated work from chat, making it easier to build and edit documents, graphics, and code. Projects grounds Claude in a team's files and past work. Albert also covers tool use, prompt generation, variables, evaluations, upcoming Claude models, and Anthropic's research into steering model features. His advice is to prototype quickly and assume model capabilities will change during product development.

## Key ideas
### AI products need to be redesigned around model capabilities
[00:01](https://www.youtube.com/watch?v=EuC1GWhQdKE&t=1s)
Albert compares the current AI moment with the early electrical industry. Factory owners first replaced steam engines with electric motors while keeping the old factory layout, so the basic inefficiencies remained. Productivity improved when factories were redesigned around electricity, with smaller specialized tools and more flexible workflows. He says many companies are doing the same thing with language models by adding a few star-icon buttons to existing products. He also compares this with early mobile software, when companies shrank websites for phones before building around cameras, GPS, and other mobile capabilities.

### Claude 3.5 Sonnet improves coding, retrieval, and multi-step work
[00:04](https://www.youtube.com/watch?v=EuC1GWhQdKE&t=4s)
Albert introduces Claude 3.5 Sonnet as the first model in Anthropic's Claude 3.5 family and says it outperforms Claude 3 Opus on several benchmarks. He is especially interested in real-world use. The model has a 200k context and strong recall for retrieval applications. In coding, he says it is better at understanding debugging problems and avoiding repeated loops. Anthropic's internal pull request evaluations gave Claude 3.5 Sonnet a 64% score, compared with 38% for Claude 3 Opus. The task requires the model to write and test its way toward a solution.

### Vision makes tables, OCR, and screenshots useful inputs
[00:05](https://www.youtube.com/watch?v=EuC1GWhQdKE&t=5s)
Albert says Claude 3.5 Sonnet improves substantially over Claude 3 Opus on Anthropic's vision benchmarks. He points to table transcription and OCR as examples that are now much easier. In the demonstration, Sonnet reproduces a table and its numbers accurately. He says vision was the most surprising capability when he began testing the model because it opens additional ways to build software. This includes working from screenshots and design files, rather than requiring every input to begin as text.

### Sonnet is cheaper and available through several cloud platforms
[00:06](https://www.youtube.com/watch?v=EuC1GWhQdKE&t=6s)
Claude 3.5 Sonnet is available through Anthropic, Amazon Bedrock, Google Vertex AI, and another platform Albert names as AWS-related in the presentation. He says developers want a choice of where they use the model. The price is $3 per million input tokens and $15 per million output tokens. Albert describes Sonnet as five times cheaper than Claude 3 Opus. He says the combination of speed, intelligence, and price makes it more practical to embed in applications.

### Artifacts turn model output into an editable working space
[00:07](https://www.youtube.com/watch?v=EuC1GWhQdKE&t=7s)
Artifacts separate Claude's produced content from the surrounding chat dialogue. Albert says people can work with Claude on essays, SVGs, and React websites. He demonstrates turning a screenshot into code and components, including a recreation of the Claude.ai chat layout in React from one screenshot. He presents this as an example of designing a product around the model's coding, reasoning, and visual abilities. The content becomes something the user can work on directly instead of remaining buried inside a conversation.

### Projects ground Claude in a team's own material
[00:08](https://www.youtube.com/watch?v=EuC1GWhQdKE&t=8s)
Projects let development teams give Claude their own knowledge, including style guides, code bases, transcripts, and earlier work. On Anthropic's team plan, people can share projects and chats with colleagues. Albert says Anthropic engineers upload repositories and documentation, then share chats and artifacts in place of Google Docs or static documentation. He presents Projects as another example of building a product experience around language models first. The model has access to the material that gives a team's work its local context.

### Developer tools now cover tool use, prompts, variables, and evaluations
[00:09](https://www.youtube.com/watch?v=EuC1GWhQdKE&t=9s)
Anthropic's tool use API lets developers give Claude client-side functions that the model can choose to call. Albert also says tool use supports consistent structured JSON output with Claude 3.5 Sonnet, and that some developers are giving Claude hundreds of tools at once. The developer console includes a prompt generator that creates a prompt from a task description. Developers can edit the generated prompt, use variables in templates, and test retrieval use cases. Anthropic is also working on an evaluation feature that was in beta in the console when he spoke.

### Product road maps should expect faster model changes
[00:10](https://www.youtube.com/watch?v=EuC1GWhQdKE&t=10s)
Albert previews Claude 3.5 Haiku and Claude 3.5 Opus and says they are planned for later that year. Anthropic wants each model generation to increase intelligence while reducing latency and cost. His advice is to build with the expectation that models will become smarter, cheaper, and faster within months rather than years. Product teams should therefore leave room to adopt a newer model during development. He also discusses interpretability research on features that activate for different topics and a possible steering API for turning specific features up or down.

### Quick prototypes create the feedback needed for AI products
[00:12](https://www.youtube.com/watch?v=EuC1GWhQdKE&t=12s)
Albert encourages developers to build quick prototypes and start validation and feedback as soon as possible. He points attendees to Anthropic's Build with Claude contest, which was running until July 10, with $10,000 in Anthropic API credits for each of the top three projects. The advice follows his broader argument that teams are still discovering what products should look like when language models are central to the workflow. Early prototypes can test whether an idea works before a team commits to a larger product build.

## Notable quotes
- "The real electrical Revolution well it didn't actually come until we imagined factories from the ground up with electricity at its core." (00:02)
- "3.5 Sonnet's combo of speed intelligence and low cost makes it much more economical to use and embed in your apps than 3 Opus." (00:07)
- "Artifacts separate the content that Claude produces from the actual chat dialogue itself." (00:07)
- "Models will become smarter cheaper and faster in orders of months not years." (00:10)
- "I want to encourage you guys to just go out there and build and make quick prototypes as fast as you can to get that validation and that feedback loop started." (00:12)

## Tools & references mentioned
- Anthropic
- Alex Albert
- Claude 3.5 Sonnet
- Claude 3 Opus
- Claude 3.5 Haiku
- Claude 3.5 Opus
- Artifacts
- Projects
- Amazon Bedrock
- Google Vertex AI
- Claude.ai
- React
- AWS
- MMLU
- HumanEval
- GPQA
- Scaling Monosemanticity
- Golden Gate Claude
- Build with Claude contest

## Who should watch
- You are adding a chat interface to an existing product and need examples of workflows designed around model output.
- Your application needs coding, retrieval, image, screenshot, tool-calling, or structured JSON capabilities from an API.
- You are planning a product over several months and want to account for lower-cost, faster, and more capable future models.

## Related talks

- [Claude Fable, Claude Tag, and Anthropic's Culture](https://aietalks.com/talks/claude-fable-claude-tag-and-anthropics-culture) (Cat Wu & Thariq Shihipar, Anthropic & Simon Willison, 51:30)
- [How Anthropic Builds: Lessons from Labs](https://aietalks.com/talks/how-anthropic-builds-lessons-from-labs) (Mike Krieger, Anthropic, 26:11)
- [Evolving Claude APIs for Agents](https://aietalks.com/talks/evolving-claude-apis-for-agents) (Katelyn Lesse, Anthropic, 13:25)
- [Keynotes & Multimodality Track](https://aietalks.com/talks/keynotes-multimodality-track) (Benjamin Duny & Simon Sturmer, AI Engineer & Chris Lattner, Modular & Auntie Aart & Mike Chambers, AWS & Alex Albert, Anthropic & Harrison Chase, LangChain & Rob Chandra, Substrate & Vic Kapati, M87 Labs & Ben Hilac, Dawn & Karan Goel, Cartesia & Chang She, LanceDB & Noah Schnapp, Character.AI & Steph Duenas & Quinn Dombrowski, Daily & Roman Huet, OpenAI & Brian Bischof, Charles Fry, Hamza Hussein, Jason Liu, Shreya Shankar & Eugene Yan & Thomas Dohmke, GitHub, 8:34:14)
- [Design like Karpathy is watching](https://aietalks.com/talks/design-like-karpathy-is-watching) (Zeke Sikelianos, Replicate, 19:26)