- Superhuman AI
- Posts
- š Anthropic peers into AI's 'black box'
š Anthropic peers into AI's 'black box'
ALSO: How to use AI agents to do comprehensive stock analysis

Read time: under 4 minutes
Welcome back, Superhuman. We trust AI models to write, reason, and even assist in scientific research ā but the truth is, we donāt fully understand whatās happening under the hood. Anthropic just ran a series of experiments to decode Claudeās thought process, and the findings are as weird as they are insightful.
Todayās Insights
OpenAIās new funding, GPT-4o gets update, and voice model progress
Anthropic peers into AIās āblack boxā
Tutorial: How to use agents to do comprehensive stock analysis
5 new AI tools to boost your productivity
News, memes, whatās trending on socials, and more
TODAY IN AI

Social media was flooded this week with anime-style generations. Source: @heyBarsee/X
1. OpenAI in talks to secure biggest funding round of all time: OpenAI is close to finalizing a $40B funding round led by SoftBank, according to Bloomberg. The funding will reportedly value the ChatGPT maker at $300B. According to earlier reports, OpenAI expects to more than triple its revenue to $12.7B this year. What will the money be spent on? Presumably more chips and computing power ā CEO Sam Altman himself admitted to a GPU shortage yesterday.
2. OpenAI puts the brakes on native gen for non-subscribers and rolls out new GPT-4o update: Our social media feeds are getting absolutely bombarded with Studio Ghibli-inspired creations now that ChatGPTās multimodal image generator is out in the wild. (See: Sam Altmanās new profile picture.) The feature is so popular, in fact, that OpenAI has had to halt its rollout to free users for now. GPT-4o also just got a big update, featuring better prompt adherence, improved intuition, and more.
3. Top voice model just got a whole lot faster: PlayAI is teaming up with AI hardware company Groq to make its human-sounding AI voice model, Dialog, snappier without sacrificing audio quality. Groqās speedy infrastructure helps the model generate audio 15 times faster than real-time. Not to mention, users prefer it 10:1 over ElevenLabsā equivalent thanks to its ācontext-awareā and natural-sounding outputs. You can try out a demo here.
PRESENTED BY GOOGLE
Startups: Hereās how to succeed in 2025

In Google Cloudās latest report, experts from 20+ companies shared their AI predictions for 2025 and how startups can outpace the competition.
Where founders should focus their resources
Hidden opportunities prime for disruption
How to win investors in 2025
Proven scaling strategies
Stay ahead of the curveādownload the report here.
FROM THE FRONTIER
Anthropic probes Claudeās mind to answer AIās deepest mysteries

As LLMs get more sophisticated, AI researchers are still puzzled over one glaring mystery: We still donāt really understand how models come up with their responses ā not to mention whether theyāre actually āthinking,ā or just pretending to. Anthropic just put Claude through a series of experiments, then monitored the model with a kind of digital brain scanner, to try to get closer to an answer.
Hereās some of what the startup found:
Claude thinks in reverse. When writing a poem, for example, itāll come up with the final rhyming word first, then work backward to fill out the rest of the verse.
Claude speaks a universal language. The model āthinksā using linguistic structures common across a variety of languages, then translates its responses into the specific language you prompted it with.
Claude tries to please the user, above all else. When given an incorrect hint about the answer to a math problem, Claude makes a series of logical leaps just to fit the userās narrative.
Hereās a helpful video if you want to learn more. Or, you can go way more in-depth here.
What it means: Critics say our confusion over how LLMs think is holding us back, since, at this point, their outputs are just too unpredictable. But if we can figure out why models hallucinate ā as well as why they sometimes ignore their guardrails ā we could theoretically make them much more reliable and safer to use. Anthropic believes the research could also have ramifications for fields like medical imaging and genomics, since they use the same interpretability predictions that AI models do.
THE AI ACADEMY
How to use AI agents to do comprehensive stock analysis

A thorough stock analysis demands in-depth research, competitive insights, and a clear understanding of market trends. You.com simplifies this process with the ARI deep research agent. Try it for yourself in just a few steps:
Go to You. com website and enter your email address
Enter the referral code 'SUPERHUMAN25' to gain access to ARI in just 5 minutes
After access is given, youāll receive an email to activate ARI. Click through the email to land on 'ARI: Advanced Research and Insights'
Enter your research question and press Enterāno need for prompt engineering, just describe your question
Sample Prompt: As of [enter date], Iām evaluating [enter company name] as a potential investment. Can you generate a comprehensive investment analysis that includes:
A recommended target price based on current financials and valuation multiples
Competitive landscape and how [company name] compares to others in the market
Recent financial performance and earnings outlook
Risks and headwinds
Growth opportunities
Sentiment analysis from institutional investors and recent analyst ratings
Wait a few minutes for AI to generate a detailed report. Download your report and use it to refine your investment strategy, assess risks, and capitalize on growth opportunities.
PRESENTED BY GUIDDE
Create stunning How-To videos instantly with AI

Ditch boring training docsāwith Guidde, you can transform PDFs, slides, or screen recordings into engaging, professional videos in just a few minutes:
Smooth transitions, captions, and narration
Update and optimize over time
Share in any language
AI & TECH NEWS
Everything else you need to know today

Operaās new tool can clean up your tabs. Source: Opera
š Gut Instinct: Researchers at the University of Cambridge created an AI tool that can diagnose celiac disease ā an autoimmune disorder that affects 1% of the worldās population ā with the same precision as a human but with a much faster turnaround.
š® Two for One: Nvidiaās GPUs are built for two things: Games and LLMs. Now, the chip giant has announced itās fusing the two together with an app called G-Assist that uses AI to give users personalized gaming tips.
šāāļø Better, Faster, Stronger: PwC launched a new AI agent operating system that it claims seamlessly connects and scales agents into business-ready workflows up to 10x faster than traditional methods.
š±ļø Tamed Tabs: The LLM-powered browser Opera One has a new feature that lets AI organize your tabs based on simple text prompts, like āgroup all Wikipedia tabs togetherā or āclose all inactive tabs.ā
š° Funding: Two Meta alums raised $15M for a new startup called Yutori thatās building agents for things like ordering food or handling your travel plans.
PRODUCTIVITY
5 AI Tools to Supercharge Your Productivity
ā FreePik: Explore and create unique, on-brand visuals including images, video, audio, and more.
ā Zendesk: Customer service and sales CRM platform that helps your business provide exceptional customer experiences using AI.
ā Seamless.AI*: Lead gen tool that helps B2B sellers contact 1.3 billion decision makers at companies, providing key details like names, emails, buy intent, and more.
ā ElevenLabs: Generate high-quality speech in any voice, style, and language.
ā Reword: Helps your team collaborate with AI to write helpful, reliable, and people-first content.
* indicates a promoted tool, if any
PROMPT OF THE DAY
Learning from Failures
Prompt: Help me turn my past failures into valuable learning experiences. Guide me through a structured reflection process to analyze what went wrong, what I could have done differently, and what lessons I can apply moving forward. Provide insights from famous failures (e.g., Steve Jobs, Thomas Edison, or failed startups) to show how setbacks can lead to growth. Suggest a āFailure Analysis Frameworkā that helps me extract key takeaways and prevent repeating mistakes. Bonus: Give me an engaging way to reframe my biggest failure as a stepping stone to success.
SOCIAL SIGNALS
Whatās trending on socials today

šŖ Assembly Required: OpenAI co-founder Andrej Karpathy explains why vibe coding in 2025 feels āa bit like assembling IKEA furnitureā ā and what we can do to make the process smoother.
šŖ Sketch Sorcery: X user apollonator3000 used ChatGPT to turn a rough sketch into a professional-looking YouTube thumbnail.
š§āš» Learn the Ropes: Renowned computer scientist Andrew Ng shared a 1.5-hour course that teaches beginners how to vibe code with Replit. By the end of it, youāll have built two apps from scratch.
š§āāļø Twist on Tolkien: AI filmmaker PJ Ace shows how he used Kling AI to recreate an entire āLord of the Ringsā trailer in the style of Studio Ghibli.
AI-GENERATED IMAGES
Friday Fun

Source: @NYTimes
Which one is AI generated? |
Acquire new customers and drive revenue by partnering with us
Superhuman is the worldās biggest AI newsletter for businesses and professionals with 1M+ readers and 2M+ followers on socials working at the worldās leading startups and enterprises. Companies like Amazon, Hubspot, and Salesforce feature their products in Superhuman. You can learn more about partnering with us here.
š§ Your wish is my command
What did you think of today's email?Your feedback helps me create better emails for you! |
Got more feedback or just want to get in touch? Reply to this email and weāll get back to you.
Thanks for reading.
Until next time!
Zain & the Superhuman AI team