- Decode AI
- Posts
- Inside OpenAIās Safety Battle
Inside OpenAIās Safety Battle
š¤ AI Models Exhibit Human-Like Number Preferences
![](https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/dbd1ff44-65dd-494d-8b18-1abd890431e8/Decode_Email_banner.gif)
Hola Decoderš
If someone forwarded this to you and you want to Decode the power of AI and be limitless, then subscribe now and Join Decode alongside 30k+ code-breakers untangling AI š§ .
OPENAI SAFETY CONCERN
š”ļø OpenAIās New Safety Team Faces Scrutiny Amid Internal Conflicts
Insights from The Verge, Tech Crunch
OpenAI has established a new safety team led by CEO Sam Altman, alongside board members Adam DāAngelo and Nicole Seligman, to oversee critical safety and security decisions. This move comes as the company faces internal challenges and external criticism about its commitment to AI safety.
The Decode:
Formation of Safety Team: The new safety team will evaluate and improve OpenAIās safety processes, presenting findings to the board, which includes Altman. Concerns persist about the lack of independent oversight, especially after key safety-focused employees like Ilya Sutskever and Jan Leike left, citing a shift away from safety priorities. Former board members Helen Toner and Tasha McCauley have also expressed doubts about OpenAIās self-governance.
I resigned a few hours before hearing the news about @ilyasut and @janleike, and I made my decision independently. I share their concerns. I also have additional and overlapping concerns.
ā Gretchen Krueger (@GretchenMarina)
10:08 PM ā¢ May 22, 2024
Internal Conflict and Altmanās Firing: Helen Toner revealed details about Sam Altmanās ousting in November 2023, citing trust issues and accusations of misconduct, including undisclosed ownership and inaccurate safety information.
The boardās decision was fueled by allegations of a toxic work environment. After pressure from employees and key investors, Altman was reinstated as CEO. Toner highlighted similar issues in Altmanās previous roles, indicating a pattern that contributed to the boardās decision.
The establishment of the new safety team and the internal conflicts surrounding Sam Altman highlight ongoing tensions within OpenAI regarding the balance between rapid AI development and stringent safety standards. With key safety personnel leaving and internal leaders managing the safety team, questions about the effectiveness and independence of these measures persist. Additionally, the internal conflict and reinstatement of Altman as CEO underscore the complexities of governance and leadership within the organization.
TOGETHER WITH MONDAY.COM
The Future of Work Management
Picture a world where workflows are finely tuned, automated to perfection, and seamlessly integrated with your favorite apps. It's not just a platform; it's a revelationāa space where managers gain unparalleled visibility into team processes, ensuring each project is a resounding success. Step into the future of work management with monday.com, where efficiency isn't a goal; it's a given.
From startups to industry giants, monday.com has transformed how teams work. Why not let your team be the next success story?
AI MODELS
š¤ AI Models Exhibit Human-Like Number Preferences
Insights from Tech Crunch
AI models often surprise us with their capabilities, but they also reveal unexpected human-like behaviors. One such behavior is their tendency to pick āfavoriteā numbers in a way that mimics human patterns.
The Decode:
Human Bias in Number Selection: Humans have a well-documented tendency to avoid picking truly random numbers. For example, people rarely choose 1 or 100, and prefer numbers like 7, avoiding patterns like multiples of 5 or repeating digits.
AIās āFavoriteā Numbers: Engineers at Gramener conducted an experiment with major LLM chatbots, asking them to pick a random number between 0 and 100. The results were surprisingly non-random, displaying distinct preferences:
OpenAIās GPT-3.5 Turbo frequently chose 47.
Anthropicās Claude 3 often picked 42.
Gemini favored 72.
Behavioral Patterns: The models avoided low and high numbers, as well as double digits and round numbers, showing biases similar to humans. For example, Claude 3 never chose numbers below 27 or above 87.
Why It Matters:
This behavior highlights that AI models, while not conscious, are trained to mimic human behavior, which includes predictable patterns even in tasks like selecting random numbers. Understanding these biases can help in refining AI training methods and managing expectations about AI behavior.
AI PREFERENCE
š Tech Firms Promote Cost-Effective, Smaller AI Models
Insights from Superhuman
Tech companies are urging businesses to use smaller, energy-efficient AI models due to their customizability and lower costs. A Financial Times analysis reveals Metaās 8 billion parameter Llama 3 model as the most economical, followed by GPT-3.5 Turbo and Gemini 1.5 Flash. In contrast, the latest standard models from OpenAI and Alphabet are the most expensive among leading AI competitors.
TOOLS YOU CANNOT MISS
ā³ TimeOS 2.0 - AI-powered New Tab page to prepare for meetings by accessing relevant context from Gmail, notes, and LinkedIn. Talk to AI directly from reminders.
š¹ Syllaby - 5-in-1 AI marketing tool for viral video creation, aiding in ideation, scheduling, scripting, and avatar-based production with an integrated video editor.
š Coverse.one - Centralize Google Docs, Sheets, Word, Excel, Figma, PDFs, and more in one place. Collaborate with your team using comments, notes, and tasks.
š Every - AI-enabled bookkeeping for startups, ensuring compliance, accurate financial statements, and understanding burn rates at a lower cost.
HOT NEWS
š Opera integrates Googleās Gemini in browsers
Source - Alex Castro / The Verge
Opera has integrated Googleās Gemini AI models into its browsers through the Aria AI extension. Aria, launched last year, acts as an AI assistant, performing tasks like answering queries and writing code allowing users to access the latest information with high performance which is available on all Opera browsers, including Opera GX.
š ļø Telegram introduces in-app Copilot bot
Source - The Verge
Microsoft has integrated an official Copilot bot into Telegram, available in beta for free on mobile and desktop. Users can search, ask questions, and have conversations with this AI chatbot. Copilot for Telegram offers features like movie recommendations, workout routines, coding assistance, translation, and internet searches. Notably, itās limited to text requests and has a 30-turn daily limit.
š§āš» Former OpenAI Researcher Joins Anthropic Over Safety Worries
Jan Leike, a former key researcher at OpenAI, has joined competitor Anthropic after resigning due to safety concerns. Jan Leike is joining Anthropic to work on scalable oversight, weak-to-strong generalization, and automated alignment research. He cited "safety concerns" as his reason for leaving OpenAI, criticizing the company for not providing enough computing resources for safety measures.
āGUARANTEEDā The only AI Tool kit youāll ever need š¤©
We know that you're a mastermind in your circle, so how about sharing the Decode with your crew?
Share our newsletter with your network, and unlock exclusive access to Decodeās ultimate AI tool book, with 500+ verified, vetted, and powerful tools across 40+ categories.
![](https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/71d98524-175c-4a37-aec4-85c27eb592bf/500%2B_Tools_GIF__1_.gif)
Just 2 referrals and you can have it all.
PS: This is not just another toolbook. Itās a handpicked compilation of the best of the best tools that actually do what they claim. Also, the list is updated with new tools every month! So get your hands on the only AI Toolbook you will ever need!
Click to copy & paste your referral link to others:
https://decode.beehiiv.com/subscribe?ref=PLACEHOLDER
Tip: Ask your friends to enter their email and then click āconfirm subscriptionā in their inbox.
You currently have 0 points of 2 points to get The Only AI Toolkit you will ever needš
Thanks for Decoding with usš„³
Your feedback is the key to our code! Help us elevate your Decode experience - hit reply and share your input on our content and style.
Keep deciphering the AI enigma, and we'll be back Tomorrow with more coded mysteries unraveled just for you!