AI that upgrades itself

Welcome back, Superhuman. I came across a viral Hacker News comment this week that puts it perfectly: It’s so easy to nitpick AI’s flaws, but the reality is that this technology would have been inconceivable even a decade ago. The fact that we’re casually debating their imperfections is, in itself, a sign of how miraculous this all is.

Today’s Insights

• Voice agents, self-improving AI, and superhuman soldiers

• Anthropic wants to help you peer into AI’s ‘black box’

• Tutorial: How to restore your old photos with AI

• 5 new AI tools to boost your productivity

• News, memes, what’s trending on socials, and more

TODAY IN AI

1. ElevenLabs’ conversational AI takes a ‘monumental step forward’: ElevenLabs, considered one of the biggest players in the voice agent space, is stepping up its game with Conversational AI 2.0. The new platform lets you build your own voice agents — including concierges, chess coaches, and game characters — with minimal latency, automatic language switching, and the ability to juggle text and voice inputs at the same time. Plus, a new turn-taking model makes it far less likely to interrupt users.

2. Startup unveils first-of-its-kind self-improving AI model: Japan’s Sakana AI is taking “self-help” to a whole new level. It just built what might be the world’s first Darwin Gödel Machine (DGM), an LLM capable of rewriting its own code to boost its performance. It can take in new information, test out hypotheses, and adjust its worldview accordingly, “just as humans do.” In one experiment, DGM jumped from 20% to 50% on the coding-focused SWE-bench by tweaking its own codebase.

3. Meta’s AI-powered headsets could give soldiers superhuman vision: The social media giant announced it’s teaming up with its former VR chief, Palmer Luckey, on futuristic helmets, glasses, and other equipment for the US Army. Luckey’s $28B defense company, Anduril, said the new AI-powered gadgets would be designed to sharpen soldiers’ senses — letting them do things like detect faraway drones or spot hidden enemies.

PRESENTED BY BOOMI

Build AI agents 65% faster (with 100% control)

Ready to build your fleet of AI agents, but afraid of losing control? 

Boomi Agentstudio is the only full agent lifecycle management solution for designing and orchestrating AI agents at scale with 100% governance. 

Trusted by AWS, Jade Global, and more: Use intuitive no-code tools to create powerful agents, enforce compliance through centralized governance, and supercharge agentic workflows across teams. 

• Monitor agent behavior in real time 

• Detect anomalies early to prevent security threats

• Register and track agents from any source with built-in visibility

See how the smartest teams scale AI agents with Boomi Agentstudio.

FROM THE FRONTIER

Anthropic shares a new way to probe AI’s thoughts

The better AI gets, the less we seem to know about it: While we’re good at tracking models’ outputs, it’s much harder to figure out the specific thought processes that led to those conclusions.

Anthropic recently got us one step closer: In March, it unveiled a new method of exploring Claude’s thoughts, zooming in to see all the “internal reasoning steps” that happened before it landed on a final answer. It uncovered some interesting findings, like the fact that Claude thinks in a “universal language” before translating to the one you’re speaking in.

Now, it’s open-sourcing the framework: That means anyone can input a small model of their choice and generate a unique graph to uncover its inner-workings. You can dig deeper by writing out a specific prompt and exploring all of the associations that pop up in the LLM’s virtual brain.

Anthropic isn’t the only one working on this: In fact, researchers from Stanford, Oxford, and other top universities just created a new benchmark that helps us measure how flattering and agreeable different models are — by tapping Reddit’s AITA subreddit as training data.

Why it matters: Visualizing models’ thought processes can help us get a better handle on why certain tweaks and edits seem to affect LLMs in different ways — and what we can do to make them more accurate and fair.

THE AI ACADEMY

How to restore your old photos with AI

• Go to Kontext Chat and sign up with your account.

• Upload your old, black-and-white image, type your prompt, and hit Enter.

Sample Prompt: ‘Add color to this image and remove all scratches‘

• Wait for a few seconds and you’ll get your old image instantly restored.

• Click on the image to open and download it.

• You can edit your images by writing simple prompts till you get your desired results.

PRESENTED BY TANKA

Messenger + Search + AI = Actually Getting Work Done

Tanka is the AI productivity powerhouse that actually understands what you want

• Turn scattered info across apps into follow-ups, work summaries, websites, product prototypes... in seconds. 

• Structured memory powers instant smart replies, internal search, and agent to execute.

• Inherent knowledge transfer lets new hires ramp up fast and cushions turnover.

Tanka generates files like pitch decks, websites, and financial forecasts automatically with information from chats and documents across your platforms.

Start delegating your work to Tanka

AI & TECH NEWS

Everything else you need to know today

🛝 AI Playground: Google released an experimental app called AI Edge Gallery that lets users browse a wide variety of open models, download them, and run them locally on their phones.

🩺 Doctor’s Orders: VC giant a16z is reportedly leading a $300M funding round for Abridge AI, whose software transcribes medical conversations so patients no longer have to rely on handwritten notes. If the deal goes through, it’ll value Abridge at $5.3B.

📱 Open for Business: Android and iOS users can now try Gemini Live, which lets you share your screen or camera and ask questions about what you see. Meanwhile, anyone can now use the new Jules coding assistant to perform 60 daily tasks at no cost.

🖼️ Next Level: Midjourney just rolled out a series of updates, including 40% faster rendering speeds and smarter image moderation.

📊 Crowded Field: A new Artificial Analysis report found that China is now neck-and-neck with the US, especially when it comes to open-source AI. And it’s not just DeepSeek competing for the top slot — there are more than 10 labs vying for the lead.

PRODUCTIVITY

5 AI Tools to Supercharge Your Productivity

✅ Tyce: Cursor for documents — craft deals, contracts, and documents in seconds.

✅ Zero: An AI-native email client that manages your inbox, so you don't have to.

✅ Peekaboo*: The AI alternative to ahrefs. See how many visitors your website gets from LLMs (vs competitors too) and how to rank higher.

✅ Ciro: Research and build the perfect prospect list in under 5 minutes with AI.

✅ Pretty Prompt: Grammarly for prompting — turn any text into working prompts.

 🎁 Want more? Check out our Top 125 AI Tools

^{* indicates a promoted tool, if any}

SOCIAL SIGNALS

What’s trending on socials today

📊 Silver Bullet: Early users are stunned by the level of detail in Perplexity Labs’ financial reports, with one former McKinsey consultant admitting the new feature “would’ve automated 70% of my job.”

📱 Engagement Hack: Speaking of Perplexity Labs, branding coach Florian Camiade just shared a prompt that lets you use the platform as a viral marketing agent. In around 10 minutes, it’ll “identify patterns of success, content strategies, engagement tactics,” and more.

🎮 Press Start: a16z’s Justine Moore shows how to turn yourself into a pixelated video game character, strolling through a setting of your choice, by combining either Flux or GPT-4o with Kling 2.1.

💭 Inner Monologue: X user Adonis Singh tried to push Claude to its limit by convincing the chatbot he was about to give away his crypto wallet code to a scammer. He discovered that Claude’s internal thoughts are very different from its actual responses.

🤖 Binary Boss: Entrepreneur Matt Shumer just appointed an AI to be the CEO of a new division of his startup. You can follow along here.

PROMPT OF THE DAY

PDCA problem solving framework

Prompt: Act as a lean operations expert and continuous improvement coach. Help me apply the PDCA (Plan-Do-Check-Act) framework to solve the following problem: [Insert the problem in detail, including background, context, and frequency]. Guide me step-by-step through the cycle.

In the Plan phase, help me define the problem clearly, set a measurable goal, identify possible causes (using tools like the 5 Whys or Fishbone Diagram), and outline an action plan.

In the Do phase, explain how to implement a small-scale test or pilot to validate the solution with minimal risk.

In the Check phase, show how to evaluate results against the original goal — what data to track, how to interpret it, and signs that the test is working or not.

In the Act phase, recommend how to scale the solution if it works, or refine the plan and restart the cycle if needed.

Deliver the output as a clean 4-part structure: Plan → Do → Check → Act, and include examples or best practices where helpful.

🤓 Want more? Check out our Top 1,000 Prompts

AI-GENERATED IMAGES

Diverse Ferry Companions

ChatGPT Prompt: A young cool woman with a tablet on the staten island ferry. She has a suitcase. Next to her is a japanese shogun, a female character from bridgeton with a big wig, Charlie Chaplin with makeup, and a female anime character with big sparkly eyes, and an alien. They have the exact matching luggage and are looking towards the woman.

Acquire new customers and drive revenue by partnering with us

Superhuman is the world’s biggest AI newsletter for businesses and professionals with 1M+ readers and 2M+ followers on socials working at the world’s leading startups and enterprises. Companies like Amazon, Hubspot, and Salesforce feature their products in Superhuman. You can learn more about partnering with us here.

🧞 Your wish is my command

Got more feedback or just want to get in touch? Reply to this email and we’ll get back to you.

Thanks for reading.

Until next time!

Zain & the Superhuman AI team