Memory Leak — #24
VC Astasia Myers’ perspectives on machine learning, cloud infrastructure, developer tools, open source, and security. Sign up here.
🚀 Products
Goat is a fine-tuned LLaMA model that significantly outperforms GPT-4 on a range of arithmetic tasks. Fine-tuned on a synthetically generated dataset, Goat achieves state-of-the-art performance on BIG-bench arithmetic sub-task. In particular, the zero-shot Goat-7B matches or even surpasses the accuracy achieved by the few-shot PaLM-540B.
Why does this matter? While general foundation models may have the ability to support more wide-ranging tasks, we are seeing that open source models fine-tuned for specific tasks can outperform them. We can imagine that companies are interested in leveraging synthetic data, LoRA, prompt-tuning, and fine-tuning to produce a model that are superior for their individual use case.
Former Github CTO Jason Warner Raises $26 Million for Foundation Model Code Startup
Jason Warner raised a $26 million seed round for an AI company called Poolside that he co-founded with serial entrepreneur Eiso Kant. Jason described GitHub’s Copilot, which assists coders with artificial intelligence, as an “interface” while describing what he hopes to build with Poolside as a “subsystem.”
Why does this matter? Poolside seems to be focused on building a foundational model around code generation. Interestingly, Jason stated Poolside is targeting a trillion parameter model. While the details of GPT-4 size are not publicly disclosed, it is estimated to be +1 trillion parameters with some believing it is 170 trillion parameters.
Commandbar released a GPT-powered chatbot for any site called Helphub AI. It syncs with any public URL or an existing help center. Users include Strava, Pitch, Evernote, and others.
Why does this matter? Chatbots have been a main use case for generative AI technologies. Teams will need to make a build versus buy decision around chatbots on their website and in their product. The number of knowledge base documents and complexity of the product the bot offers Q&A for will be determining factors when considering to build a chatbot.
📰 Content
Snowflake announced that it has agreed to acquire Neeva, a search startup founded by former Google executives. The deal, announced during Snowflake’s quarterly earnings report, is expected to enhance Snowflake’s ability to offer intelligent and conversational search experiences to its customers who use its platform to store, analyze and share data.
Why does this matter? This acquisition will enable Snowflake users and application developers to build rich search-enabled and conversational experiences. In March 2022 Snowflake acquired Streamlit for $800M to broaden its user scope to include data scientists and machine learning engineers and enable the creation of data apps on Snowflake. Now Snowflake is seeing that with generative AI software engineers and non-technical users are able to get value out of data and create AI applications. Neeva will help them reach these two user bases to further Snowflake usage across the organization and move further up the stack.
3 Things Businesses Must Do to Secure Applications in the AI Era
Reed McGinley-Stempel, the co-founder and CEO at Stytch, which offers authentication products, states there are new threats fueled by AI: 1) more sophisticated bot traffic, 2) more believable phishing attacks, and 3) the rise of legitimate AI agents accessing customers’ online accounts on behalf of users.
Why does this matter? Each security threat that Reed states is coming to fruition. AI is being used to mimic humans in order to fool humans. AI agents accessing accounts on behalf of users is a particularly interesting threat vector as AutoGPT, BabyAGI, Cognosis, and AgentGPT operate in the browser to perform tasks. It is important to regularly test the security of AI systems to identify vulnerabilities.
A Majority of Americans Have Heard of ChatGPT, but Few Have Tried It Themselves
About six-in-ten U.S. adults (58%) are familiar with ChatGPT, though relatively few (14%) have tried it themselves, according to a Pew Research Center survey conducted in March 2023. Among those who have tried ChatGPT, a majority report it has been at least somewhat useful. Those with higher household incomes and more formal education are more likely to know about ChatGPT.
Why does this matter? While the survey is about two months old, it reinforces technical awareness and understanding varies. We can imagine that the percentage of U.S. adults familiar with ChatGPT is now closer to 75%, and those that have tried it are closer to 33%. After all, South Park had an episode that highlighted ChatGPT.
💼 Jobs
⭐️Claypot — Founding Engineer (Infra)
⭐️Chroma — Member of Technical Staff
⭐️Speakeasy — Founding Engineer
