|
4 months in the past, we launched Anthropic’s Claude 3.5 in Amazon Bedrock, elevating the trade bar for AI mannequin intelligence whereas sustaining the pace and value of Claude 3 Sonnet.
Right now, I’m excited to announce three new capabilities for the Claude 3.5 mannequin household in Amazon Bedrock:
Upgraded Claude 3.5 Sonnet – You now have entry to an upgraded Claude 3.5 Sonnet mannequin that builds upon its predecessor’s strengths, providing much more intelligence on the similar price. Claude 3.5 Sonnet continues to enhance its functionality to resolve real-world software program engineering duties and comply with advanced, agentic workflows. The upgraded Claude 3.5 Sonnet helps throughout the whole software program improvement lifecycle, from preliminary design to bug fixes, upkeep, and optimizations. With these capabilities, the upgraded Claude 3.5 Sonnet mannequin might help construct extra superior chatbots with a heat, human-like tone. Different use circumstances by which the upgraded mannequin excels embody information Q&A platforms, knowledge extraction from visuals like charts and diagrams, and automation of repetitive duties and operations.
Pc use – Claude 3.5 Sonnet now gives pc use capabilities in Amazon Bedrock in public beta, permitting Claude to understand and work together with pc interfaces. Builders can direct Claude to make use of computer systems the way in which individuals do: by a display, shifting a cursor, clicking buttons, and typing textual content. This works by giving the mannequin entry to built-in instruments that may return pc actions, like keystrokes and mouse clicks, enhancing textual content information, and working shell instructions. Software program builders can combine pc use of their options by constructing an action-execution layer and grant display entry to Claude 3.5 Sonnet. On this manner, software program builders can construct purposes with the flexibility to carry out pc actions, comply with a number of steps, and test their outcomes. Pc use opens new prospects for AI-powered purposes. For instance, it may possibly assist automate software program testing and again workplace duties and implement extra superior software program assistants that may work together with purposes. Given this expertise is early, builders are inspired to discover lower-risk duties and use it in a sandbox setting.
Claude 3.5 Haiku – The brand new Claude 3.5 Haiku is coming quickly and combines speedy response instances with improved reasoning capabilities, making it splendid for duties that require each pace and intelligence. Claude 3.5 Haiku improves on its predecessor and matches the efficiency of Claude 3 Opus (beforehand Claude’s largest mannequin) on the pace and value of Claude 3 Haiku. Claude 3.5 Haiku might help with use circumstances equivalent to quick and correct code recommendations, extremely interactive chatbots that want speedy response instances for customer support, e-commerce options, and academic platforms. For patrons coping with giant volumes of unstructured knowledge in finance, healthcare, analysis, and extra, Claude 3.5 Haiku might help effectively course of and categorize data.
In response to Anthropic, the upgraded Claude 3.5 Sonnet delivers across-the-board enhancements over its predecessor, with important features in coding, an space the place it already excelled. The upgraded Claude 3.5 Sonnet exhibits wide-ranging enhancements on trade benchmarks. On coding, it improves efficiency on SWE-bench Verified from 33% to 49%, scoring increased than all publicly accessible fashions. It additionally improves efficiency on TAU-bench, an agentic instrument use process, from 62.6% to 69.2% within the retail area, and from 36.0% to 46.0% within the airline area. The next desk consists of the mannequin evaluations supplied by Anthropic.
Pc use, a brand new frontier in AI interplay
As a substitute of limiting the mannequin to make use of APIs, Claude has been skilled on common pc abilities, permitting it to make use of a variety of ordinary instruments and software program applications. On this manner, purposes can use Claude to understand and work together with pc interfaces. Software program builders can combine this API to allow Claude to translate prompts (for instance, “discover me a resort in Rome”) into particular pc instructions (open a browser, navigate this web site, and so forth).
Extra particularly, when invoking the mannequin, software program builders now have entry to 3 new built-in instruments that present a digital set of fingers to function a pc:
- Pc instrument – This instrument can obtain as enter a screenshot and a aim and returns an outline of the mouse and keyboard actions that needs to be carried out to attain that aim. For instance, this instrument can ask to maneuver the cursor to a selected place, click on, kind, and take screenshots.
- Textual content editor instrument – Utilizing this instrument, the mannequin can ask to carry out operations like viewing file contents, creating new information, changing textual content, and undoing edits.
- Bash instrument – This instrument returns instructions that may be run on a pc system to work together at a decrease degree as a consumer typing in a terminal.
These instruments open up a world of prospects for automating advanced duties, from knowledge evaluation and software program testing to content material creation and system administration. Think about an utility powered by Claude 3.5 Sonnet interacting with the pc simply as a human would, navigating via a number of desktop instruments together with terminals, textual content editors, web browsers, and in addition able to filling out kinds and even debugging code.
We’re excited to assist software program builders discover these new capabilities with Amazon Bedrock. We count on this functionality to enhance quickly within the coming months, and Claude’s present means to make use of computer systems has limits. Some actions equivalent to scrolling, dragging, or zooming can current challenges for Claude, and we encourage you to start out exploring low-risk duties.
When OSWorld, a benchmark for multimodal brokers in actual pc environments, the upgraded Claude 3.5 Sonnet at the moment will get 14.9%. Whereas human-level ability is way forward with about 70-75%, this result’s significantly better than the 7.7% obtained by the next-best mannequin in the identical class.
Utilizing the upgraded Claude 3.5 Sonnet within the Amazon Bedrock console
To get began with the upgraded Claude 3.5 Sonnet, I navigate to the Amazon Bedrock console and select Mannequin entry within the navigation pane. There, I request entry for the brand new Claude 3.5 Sonnet V2 mannequin.
To check the brand new imaginative and prescient functionality, I open one other browser tab and obtain from the Our World in Information web site the Wind energy technology chart in PNG format.
Again within the Amazon Bedrock console, I select Chat/textual content underneath Playgrounds within the navigation pane. For the mannequin, I choose Anthropic because the mannequin supplier after which Claude 3.5 Sonnet V2.
I exploit the three vertical dots within the enter part of the chat to add the picture file from my pc. Then I enter this immediate:
That are the highest international locations for wind energy technology? Reply solely in JSON.
The end result follows my directions and returns the record extracting the data from the picture.
Utilizing the upgraded Claude 3.5 Sonnet with AWS CLI and SDKs
Right here’s a pattern AWS Command Line Interface (AWS CLI) command utilizing the Amazon Bedrock Converse API. I exploit the --query parameter of the CLI to filter the end result and solely present the textual content content material of the output message:
In output, I get this textual content within the response.
An anchor! You throw an anchor out once you need to use it to cease a ship, however you're taking it in (pull it up) when you do not need to use it and need to transfer the boat.
The AWS SDKs implement the same interface. For instance, you should utilize the AWS SDK for Python (Boto3) to investigate the identical picture as within the console instance:
import boto3
MODEL_ID = "anthropic.claude-3-5-sonnet-20241022-v2:0"
IMAGE_NAME = "wind-generation.png"
bedrock_runtime = boto3.shopper("bedrock-runtime")
with open(IMAGE_NAME, "rb") as f:
picture = f.learn()
user_message = "That are the highest international locations for wind energy technology? Reply solely in JSON."
messages = [
{
"role": "user",
"content": [
{"image": {"format": "png", "source": {"bytes": image}}},
{"text": user_message},
],
}
]
response = bedrock_runtime.converse(
modelId=MODEL_ID,
messages=messages,
)
response_text = response["output"]["message"]["content"][0]["text"]
print(response_text)
Integrating pc use along with your utility
Let’s see how pc use works in observe. First, I take a snapshot of the desktop of a Ubuntu system:
This screenshot is the place to begin for the steps that will likely be applied by pc use. To see how that works, I run a Python script passing in enter to the mannequin the screenshot picture and this immediate:
Discover me a resort in Rome.
This script invokes the upgraded Claude 3.5 Sonnet in Amazon Bedrock utilizing the brand new syntax required for pc use:
import base64
import json
import boto3
MODEL_ID = "anthropic.claude-3-5-sonnet-20241022-v2:0"
IMAGE_NAME = "ubuntu-screenshot.png"
bedrock_runtime = boto3.shopper(
"bedrock-runtime",
region_name="us-east-1",
)
with open(IMAGE_NAME, "rb") as f:
picture = f.learn()
image_base64 = base64.b64encode(picture).decode("utf-8")
immediate = "Discover me a resort in Rome."
physique = {
"anthropic_version": "bedrock-2023-05-31",
"max_tokens": 512,
"temperature": 0.5,
"messages": [
{
"role": "user",
"content": [
{"type": "text", "text": prompt},
{
"type": "image",
"source": {
"type": "base64",
"media_type": "image/jpeg",
"data": image_base64,
},
},
],
}
],
"instruments": [
{ # new
"type": "computer_20241022", # literal / constant
"name": "computer", # literal / constant
"display_height_px": 1280, # min=1, no max
"display_width_px": 800, # min=1, no max
"display_number": 0 # min=0, max=N, default=None
},
{ # new
"type": "bash_20241022", # literal / constant
"name": "bash", # literal / constant
},
{ # new
"type": "text_editor_20241022", # literal / constant
"name": "str_replace_editor", # literal / constant
}
],
"anthropic_beta": ["computer-use-2024-10-22"],
}
# Convert the native request to JSON.
request = json.dumps(physique)
attempt:
# Invoke the mannequin with the request.
response = bedrock_runtime.invoke_model(modelId=MODEL_ID, physique=request)
besides Exception as e:
print(f"ERROR: {e}")
exit(1)
# Decode the response physique.
model_response = json.hundreds(response["body"].learn())
print(model_response)
The physique of the request consists of new choices:
anthropic_betawith worth["computer-use-2024-10-22"]to allow pc use.- The
instrumentspart helps a brand newkindchoice (set tocustomizedfor the instruments you configure). - Observe that the pc instrument must know the decision of the display (
display_height_pxanddisplay_width_px).
To comply with my directions with pc use, the mannequin gives actions that function on the desktop described by the enter screenshot.
The response from the mannequin features a tool_use part from the pc instrument that gives step one. The mannequin has discovered within the screenshot the Firefox browser icon and the place of the mouse arrow. Due to that, it now asks to maneuver the mouse to particular coordinates to start out the browser.
{
"id": "msg_bdrk_01WjPCKnd2LCvVeiV6wJ4mm3",
"kind": "message",
"function": "assistant",
"mannequin": "claude-3-5-sonnet-20241022",
"content material": [
{
"type": "text",
"text": "I'll help you search for a hotel in Rome. I see Firefox browser on the desktop, so I'll use that to access a travel website.",
},
{
"type": "tool_use",
"id": "toolu_bdrk_01CgfQ2bmQsPFMaqxXtYuyiJ",
"name": "computer",
"input": {"action": "mouse_move", "coordinate": [35, 65]},
},
],
"stop_reason": "tool_use",
"stop_sequence": None,
"utilization": {"input_tokens": 3443, "output_tokens": 106},
}That is simply step one. As with traditional instrument use requests, the script ought to reply with the results of utilizing the instrument (shifting the mouse on this case). Primarily based on the preliminary request to e-book a resort, there could be a loop of instrument use interactions that can ask to click on on the icon, kind a URL within the browser, and so forth till the resort has been booked.
A extra full instance is obtainable on this repository shared by Anthropic.
Issues to know
The upgraded Claude 3.5 Sonnet is obtainable at the moment in Amazon Bedrock within the US West (Oregon) AWS Area and is obtainable on the similar price as the unique Claude 3.5 Sonnet. For up-to-date data on regional availability, check with the Amazon Bedrock documentation. For detailed price data for every Claude mannequin, go to the Amazon Bedrock pricing web page.
Along with the better intelligence of the upgraded mannequin, software program builders can now combine pc use (accessible in public beta) of their purposes to automate advanced desktop workflows, improve software program testing processes, and create extra refined AI-powered purposes.
Claude 3.5 Haiku will likely be launched within the coming weeks, initially as a text-only mannequin and later with picture enter.
You’ll be able to see how pc use might help with coding on this video with Alex Albert, Head of Developer Relations at Anthropic.
This different video describes pc use for automating operations.
To study extra about these new options, go to the Claude fashions part of the Amazon Bedrock documentation. Give the upgraded Claude 3.5 Sonnet a attempt within the Amazon Bedrock console at the moment, and ship suggestions to AWS re:Submit for Amazon Bedrock. You could find deep-dive technical content material and uncover how our Builder communities are utilizing Amazon Bedrock at group.aws. Tell us what you construct with these new capabilities!
– Danilo