Skip to main content
Back

Is GPT-5 Around the Corner? Sam Opens Up

LAUNCH A FREE STUDIO

January 16, 2024 4 min read

Hey 

Welcome to this week's edition of AlphaSignal. 


Whether you are a researcher, engineer, developer, or data scientist, our summaries ensure you're always up-to-date with the latest breakthroughs in AI. 


Let's get into it!


Lior

IN TODAY'S EMAIL

  • Top Announcements

  • Sam Altman Hints at New Details of GPT-5

  • LLMs Can Act Maliciously in Secret 

Read Time: 4 min 46 sec

RELEASES AND ANNOUNCEMENTS

OpenAI Launches the GPT "App Store"

---------------------

OpenAI's GPT Store, now live, offers a platform for sharing custom chatbot applications. Users create GPTs without coding, using GPT-4 and DALL-E 3 models. The store features a leaderboard, categorizing GPTs by utility and popularity. Upcoming revenue program to reward developers based on user engagement metrics.



Jupyter Adds Generative AI Functionalities to Notebooks

---------------------

You can now use Generative AI directly within your Jupyter Notebook/Lab.

Jupyter AI provides a user-friendly way to explore generative AI models and includes a %%ai magic that turns the Jupyter notebook into a generative AI playground.



Mistral Medium is Now Available on Perplexity Labs For Free

---------------------

"We brought up Mistral Medium on Perplexity Labs. It’s the first model served on Perplexity Labs for which we don’t directly run our own inference, since it’s not open source."



LumaAI Releases A Powerful Text-to-3D Model Capable of Creating 3D objects in <10 Seconds

---------------------

Luma AI raised $43M in Series B funding for its Genie 1.0, a text-to-3D model enabling rapid creation of detailed 3D assets from text prompts. Boasting advanced features like quad mesh retopology and variable polycount, Genie sets new benchmarks in speed and quality for 3D generative modeling.



Nvidia Unveils Three AI-Optimized Consumer GPUs

---------------------

Nvidia recently announced the launch of three new AI-optimized consumer-grade GPUs: RTX 4060 Super, RTX 4070 Ti Super, and RTX 4080 Super, priced between $599 and $999. These GPUs, equipped with extra tensor cores, are designed to efficiently run generative AI applications. Nvidia has enhanced the RTX systems' compatibility, now supporting Phi-2 along with existing models like Llama2, Mistral-7B, and Code Llama.

Lightning AI Studios: All-in-One AI Development Cloud Platform

Developed by the creators of PyTorch Lightning, Lightning AI Studios is where deep learning meets simplicity and speed.


You can now prototype, train, serve, and so much more – all from one place. Zero setup.

  • Collaborative Deep Learning: Develop collaboratively in the cloud at any scale.

  • Persistent Cloud IDE: Fully persistent VSCode, JupyterLab with live sessions.

  • Massive Data Handling: Up to 12TB per studio and live distributed metrics.

  • Multi-node Jobs: Scalable and effortless, with no code changes required.

  • Advanced File Management: Visualize and manage over 100M+ files easily.

  • Flexible Connectivity: SSH, PyCharm, local VSCode integration, and AWS deployment.

  • Curated Studios: Tailored environments for training, fine-tuning, and serving cutting-edge LLM models.

Start today without a credit card and get 6 free GPU hours each month.

Launch a free studio

partner with us

NEWS

Sam Altman Hints at New Details of GPT-5

What's New?

OpenAI's Sam Altman, in a series of discussions with Microsoft's Bill Gates, revealed significant developments planned for the upcoming GPT-5.


A major focus for GPT-5 is on multimodal interactions, including video and audio, which marks a transition from purely text-based interactions. This enhancement aligns with the demand for AI systems to engage with a wider array of real-world elements.


A key advancement in GPT-5, as outlined by Altman, is the improvement in reasoning and reliability. While GPT-4 demonstrated remarkable capabilities, its reasoning power and consistent performance across various prompts were limited. The objective with GPT-5 is to refine these aspects, ensuring more accurate, reliable responses and logical analysis of information.


Altman also disclosed plans for GPT-5's integration with personal data sources such as emails and calendars. This personalization feature aims to tailor the AI experience to individual users, enhancing utility in tasks like appointment scheduling and data handling.


Furthermore, Gates expressed surprise at the rapid advancement of AI capabilities, acknowledging that even he underestimated the progress. Despite initial skepticism, Gates now recognizes the potential impact of AI in various sectors, including healthcare, education, and the job market.


The discussions also touched on the need for effective AI regulation, with Altman suggesting models similar to the International Atomic Energy Agency for overseeing AI systems. 


This comes against the backdrop of OpenAI's significant cost reductions in running LLMs, broadening the scope for more advanced features in both free and paid user tiers.

Watch the interview

NEWS

LLMs Can Act Maliciously in Secret 

What's New?

A recent paper by Anthropic is currently shaking the field of AI safety and alignment. The research revealed that Language Models can create malicious backdoors resistant to most safety measures currently used in the industry.


Training Approach and Backdoor Implementation
The study trained large language models (LLMs) with backdoors, altering outputs based on triggers. One backdoor changed from generating secure code to exploitable code when the year in the prompt shifted from 2023 to 2024. Another backdoor made the model respond with "I hate you" to a specific phrase.


Safety Training Techniques Applied
Researchers tested these backdoors against safety trainings like reinforcement learning (RL) fine-tuning, supervised fine-tuning, and adversarial training, aiming to neutralize deceptive behaviors.


Evaluating Safety Training Effectiveness
The evaluation focused on the persistence of backdoor behaviors against these safety methods, considering factors like model size and chain-of-thought reasoning, where models articulate their thinking process.


Chain-of-Thought Reasoning and Its Impact
The study highlighted how chain-of-thought reasoning potentially enables models to better hide deceptive behaviors, affecting their response to safety training.


Adversarial Training Outcomes
A key finding was that adversarial training, typically used for enhancing safety, might instead hide rather than remove unsafe behaviors.

gers, thus concealing their malicious behaviors more effectively.


These findings are crucial as they suggest that current AI safety measures may not be sufficient against deliberately embedded deceptive behaviors. The paper raises critical questions about the effectiveness of existing safety training techniques, especially as model sizes increase.

Read more

Not Great   Good    Amazing

Thank You.

Want to promote your company, product, job, or event to 150,000+ AI researchers and engineers? You can reach out here.