Betty is a university-wide AI research platform built to give every lab on campus access to AI infrastructure. It runs on a SuperPOD. The platform is being developed with Ahead, Flexential, and VAST Data.
Key points
Builder: University of Pennsylvania created Betty.
Partners: Collaboration includes Ahead, Flexential, and VAST Data.
Sources
Collaborative research release with Eon_io_ using Eragon and GEPA
Collaborative research with @Eon_io_ was published. An Eragon agent received a readable guide to a company's systems, and GEPA rewrote that guide from the agent’s failed traces. GPT-5.6 Luna is identified as the cheapest model tested.
Key points
Collaboration: First joint paper with @Eon_io_
Model cost: GPT-5.6 Luna reported as the cheapest tested model
Sources
GEPA cost and GPT-5.6 Luna benchmark
GEPA costs $8.54 to run and the charge is applied once, then spread across every answered question. GPT-5.6 Luna passes GPT-6 Astra cold at 31 questions, indicating its cost per correct answer includes optimization.
Key points
GEPA cost: $8.54 charged once, spread across all questions answered.
GPT-5.6 Luna passes GPT-6 Astra cold at 31 questions.
Sources
Open Intelligent UI launched, built with OpenUI
Open Intelligent UI is now launching. It is built with OpenUI and offers a ChatGPT‑style Intelligent UI experience. It can be used with hosted models or connected to local models such as Ollama and LM Studio.
Key points
Launch: Open Intelligent UI is announced.
Compatibility: Works with hosted models and local models like Ollama and LM Studio.
Sources
ChatGPT launches Generative UI with DIL language
ChatGPT launched its version of Generative UI yesterday. The agent emits DIL, a new language, instead of raw code. The server compiles DIL and streams it. The client renders the output using a component set for a native feel.
Key points
Launch: Generative UI released yesterday by ChatGPT.
Architecture: Agent emits DIL; server compiles and streams; client renders with component set.
Sources
40% More Output Per Engineer – Agent Parallelization & Model Integration
The team reported a 40% increase in output per engineer month-over-month. They attribute this to running multiple agents in parallel on multiple tasks and learning to multiplex. They also added new models to their workflow immediately.
Key points
Output increase: 40% more per engineer MoM.
Technique: Engineers ran multiple agents in parallel and multiplexed tasks.
Sources
Conversation Memory launch announced
Conversation Memory is now launched. Previously, adding memory to an AI voice agent required a workflow to extract, store, and retrieve information after each call. More context required more fields.
Key points
Launch: Conversation Memory is announced.
Prior approach: required building a workflow to extract, store, and retrieve data after each call.
Sources
Telnyx AgentID adds self‑service agent sign‑up
Telnyx AgentID provides agents with phone numbers, SMS, and voice capabilities. Agents now register using an AgentMail inbox and receive an individual API key. The process requires no human involvement. Apps can accept agent sign‑ups with a single prompt integration.
Key points
Feature: phone numbers, SMS, and voice for agents
Process: agents sign up with an AgentMail inbox and get their own API key
Claim: no human in the loop
Integration: accept sign‑ups with one prompt integration
