In partnership with

Anthropic just quadrupled prices on its smallest model. Google races to release increasingly powerful versions of Gemini. As major tech companies double down on model development open source technology continues to advance.

The vanishing lead?

Open source models now trail their closed counterparts by just 5-22 months in capability. This shrinking gap could explain the frenetic pace of investment and development from incumbent players. When Meta released Llama, it triggered an explosion of innovation that proved impossible to contain. Now, even Google's past internal memos acknowledge they perhaps lack effective defenses against this open-source momentum.

Beyond raw performance 

The battle isn't just about processing power and parameter counts. Recent MIT research reveals that large language models fundamentally lack a coherent understanding of the world, suggesting we may need entirely new approaches. This creates an interesting dynamic: while commercial players pour resources into scaling existing architectures, open collaboration might actually be better suited to discover breakthrough approaches.

Looking forward 

Tech giants' large scale investments suggest they see this as an existential moment - and they may be right. If open-source models continue their rapid progress, we could see a fundamental restructuring of power in the AI industry. Will the AI ecosystem be dominated by a few powerful players or will open source be the backbone of the tech stack?

We publish daily research, playbooks, and deep industry data breakdowns. Learn More Here

The Bagel team just published new research on two complementary methods for advancing AI reasoning: training-time and inference-time techniques.

Training-Time Enhancements involve refining model structures. Methods like Parameters Efficient Fine-Tuning (PEFT) optimize learning efficiency by targeting specific neural pathways within fixed frameworks, while approaches like WizardMath’s 3-step reasoning improve structured reasoning.

Inference-Time Enhancements focus on on-the-fly thinking, using Chain of Thought prompting to activate logical processing without extra training. Techniques like Self-Consistency for validation and Program of Thought for coding tasks enable precise, stepwise reasoning.

The findings suggest combining both approaches: training-time builds foundational reasoning, while inference-time optimizes its application. Bagel Network supports this advancement through open-source infrastructure, empowering a community-driven AI evolution though monetizable open source AI.

About Bagel: Bagel is an AI & cryptography research lab, building the world's first monetization layer for open-source AI.

In this episode, the hosts Anthony Batt and Shane Robinson with guest Joe Veroneau from Conveyor discuss outsmarting paperwork. Conveyor is a company that helps automate security reviews and document sharing between companies. They use AI technology, specifically language models, to automate the process of filling out security questionnaires. This saves customers a significant amount of time and improves the quality of their responses.