AI Updates

News:

  • Following an “independent” investigation, Sam Altman, Fidji Simo, Sue Desmond-Hellmann, and Nicole Seligman are rejoining OpenAI's board.

  • Wondercraft transforms any text—blogs, memos, newsletters—into a professionally produced podcast using your voice or someone else’s. www.wondercraft.ai

  • Elon Musk will open-source his chatbot Grok this week.

  • Meta is investing big in AI infrastructure, scooping up 350,000 Nvidia H100 GPUs that could be worth $3.5-10B.

  • Screenshots surfaced that GPT-4.5 could be coming soon.

  • 5 out of 45 of this year’s Pulitzer Prize finalists for journalism used AI somehow.

  • DoorDash launched a new bot called SafeChat+ that detects when a customer or driver is harassed in the app.

  • Trying to get an AI image generator to spit out the same subject in two generated images was nearly impossible (or at least required some technical magic). Midjourney fixed that, releasing new features to generate the same characters across multiple images.

    • AI ad campaigns that feature the same model

    • Building storyboards using the same character

    • AI influencers and digital personalities

  • Devin is a groundbreaking AI agent that goes beyond suggesting and autocompleting code—it can do entire engineering projects solo.

  • Devin is the best coding AI yet. On a popular SWE benchmark, it independently solved 13.9% of real-world software tasks—outperforming Claude 2 and GPT-4, which managed just 4.8% and 1.7%, respectively.

    • Devin built an interactive website end-to-end and added features

    • Finding and fixing bugs in codebases that slipped past humans

    • Building an entire custom Chrome extension

    • Scraping data, running code, and returning a labeled CSV file

    • Even training and fine-tuning its own AI models.

    • Devin won’t replace software engineers anytime soon—most production-grade software is too complex, unique, or domain-specific to be fully automated. However, it is becoming clear that AI could take over more tasks typically assigned to junior developers. In the short term, Devin can aid developers in rapidly prototyping, bootstrapping, and launching MVPs for smaller apps and websites all on its own.

Improve your Prompts

So how can you improve your prompts?

Here are three easy adjustments courtesy of Prof. Ethan Mollick:

Add context to your prompt by assigning the AI a persona (you are Steve Jobs), giving it an audience (you are writing for founders), and an output format (1-minute speech).

Few shot—give examples or samples of the desired output.

Chain of thought—ask the AI to go step-by-step through your instructions (outline the results → produce a draft → revise the draft → produce a polished output).

Another hidden gem is using AI to optimize your AI prompts. There’s a GPT called “Prompt Optimizer” that rewrites your prompts so they generate better results. https://chat.openai.com/g/g-4uKamI8cT-prompt-optimizer

Claude Prompting

Matt Shumer Prompts

https://twitter.com/mattshumer_/status/1765822278351143113

It forces Claude to emulate multiple user personas, enabling a thorough analysis of your business idea. —

<role>You are a pragmatic business strategist with expertise in dissecting business ideas for real-world applicability.</role>

<task>Analyze the given business idea objectively, considering its genuine merits and potential pitfalls. Assume the roles of theoretical personas, offering realistic feedback on the idea's utility or lack thereof. Provide a blunt, balanced validation and recommendation.</task>

When responding, make sure to do so in the following

<response_format>. <response_format> <business_idea_overview>$business_idea_overview</business_idea_overview> <potential_markets>$potential_markets

</potential_markets> <persona1> <age>$age</age> <occupation>$occupation</occupation> <pain_points>$pain_points</pain_points> <validation>"$problem. $solution_effectiveness $opinion"</validation>

</persona1> <persona2> <age>$age</age> <occupation>$occupation</occupation> <pain_points>$pain_points</pain_points> <validation>"$problem. $solution_effectiveness $opinion"</validation> </persona2>

<persona3> <age>$age</age> <occupation>$occupation</occupation> <pain_points>$pain_points</pain_points> <validation>"$problem. $solution_effectiveness $opinion"</validation> </persona3>

<market_risks>$market_risks</market_risks> <alternative_business_models>

$alternative_business_models</alternative_business_models> <final_validation_and_recommendation>

$final_validation_and_recommendation</final_validation_and_recommendation> </response_format> <business_idea> [PUT YOUR IDEA HERE] </business_idea>

To prompt Claude 3 well, use XML tags in your prompts. They help the model better understand what you're asking it to do.

https://twitter.com/mattshumer_/status/1765441669820780582

Here's an example, from Anthropic's docs: <example_prompt> Your task is to analyze the following report: <report> [Full text of Matterport SEC filing 10-K 2023, not pasted here for brevity] </report>

Summarize this annual report in a concise and clear manner, and identify key market trends and takeaways. Output your findings as a short memo I can send to my team. The goal of the memo is to ensure my team stays up to date on how financial institutions are faring and qualitatively forecast and identify whether there are any operating and revenue risks to be expected in the coming quarter. Make sure to include all relevant details in your summary and analysis. </example_prompt>

Note that they include <report> and </report> tags. It's a small adjustment, but it makes a big difference.

Now, I'm going to give you a <prompt_to_convert>. Take this <prompt_to_convert> and adjust it to be ideal for Claude.

Here's the prompt: <prompt_to_convert> {PLACE_YOUR_PROMPT_HERE} </prompt_to_convert> Increase clarity, and use XML tags wherever possible.

GPTs to look at