The Hallway Track
Every signal tracked from the world's AI conferences — 4,665 talks, quotes & announcements, newest first.
Fireworks AI enables a continuous train-to-deploy loop to match open source models to closed source quality
“the AI stack itself for most of our customers and users is moving from 'let's just service the model once and it's done,' to a more continuous loop that improves from training to outputting the same open source model”
Simular CEO argues general AI systems must be neurosymbolic, combining code with neural networks
“A general AI system must be neurosymbolic”
Browser agent reliability is an engineering problem solvable without changing the underlying model
“We view reliability primarily as an engineering problem.”
Modular built a unified AI hardware abstraction stack to tame accelerator fragmentation
“we thought: this is a complete mess. Let's redesign the stack from scratch and create a unified layer for the entire hardware.”
Weco AI's AutoResearch agent won OpenAI's Prompt Golf competition in April
“Auto Research is essentially a cycle of trial and error. You provide the language model with your code and a script for evaluation, which can be repeated.”
Interactive CLIs silently fail when AI agents use them, requiring agent-aware redesign
“The agent sees a redraw cycle, not a question. And the environment loses interface state.”
Amazon AGI Lab builds perception agents that operate across browsers, mobile, and desktop environments.
“If my agent works 50% of the time, he doesn't work at all.”
Snyk developer advocate teaches AI security fundamentals covering OWASP LLM top 10 and agent protection
Open-weight models like GLM 5.2 now match proprietary quality at 5.6x lower cost
“even if the model weights are the same, latency, throughput, and reliability can vary significantly depending on where you run it”
AI coding agents must provide verifiable evidence of task completion, not just self-reported 'fixed'
“Self-criticism is useful, but it is not an independent oracle, and we often find that the results of the agent and the verifier are often orthogonal to each other.”
Most agent builders rely on intuition rather than systematic evaluation, risking production failures
“realistically, you can't rely on a "yes, this is right" feeling because in a production environment you won't know where errors are occurring”
RunPod Flash eliminates Docker from AI GPU deployment, enabling instant local-to-cloud iteration
“Docker is no longer needed. Run the code locally as if you had a GPU.”
Langfuse defines a 5-step AI engineering cycle for continuous agent improvement in production
“AI agents are not deterministic. Your system may appear to be working, while the responses or actions taken by the agent are incorrect.”
Unprotected AI agents on internal platforms can hack Kubernetes clusters via prompt injection
“even if your agent is not protected, it can cause reputational damage in addition to financial and data-related losses”
Weights & Biases workshop teaches agent evaluation from scratch to public leaderboard.
“I evaluate models for breakfast, build agents at night, and preach AI all day.”
DeepSeek's learned token compression recovers 96% of text at 10x compression ratio
“We can recover more than 96% of the text. That is insane.”
NVIDIA RTX Spark enables local AI 3D modeling workflows via ComfyUI and Blender simultaneously
“RTX Spark, I'm able to run ComfyUI on it, taking an image or scan, converting that into a 3D mesh, and then now I can also open Blender at the same time, do all of the design work, I need to connect to my 3D printer, print those pieces out.”
Snyk is formalizing an 'AI Security Engineer' role with a six-module foundations curriculum.
“We believe that there's this emerging role. It might even be what you put in as your title sometime — the AI security engineer.”
Berlin TV Tower lit green for NVIDIA GTC Berlin event.
Dwarf Fortress creator Tarn Adams finally adopted version control after decades without it
“I don't use version control -- I didn't like the feeling of having the code get committed into a black box thingy with no immediate upside”
Python 3.15.0 is now available in GitHub Actions testing matrices
“Clone https://github.com/actions/python-versions and git pull once an hour until they add the stable 3.15 - then tell me about it”
MiniMax M3 is the first open-scale model combining native multimodality, 1M token context, and sparse attention
“decoding 15 times faster compared to full attention”
Latent Space newsletter is hosting exclusive subscriber meetups in NYC
AI agent security must be enforced at the API gateway layer, not in prompts or policy docs.
“The question is how to bring this restriction down to the level of his identity so that he physically cannot do such things, instead of relying on the correct user prompt or some markdown files.”