X / Twitter
Matrix3D: Large Photogrammetry Model All-in-One (with code)
TL;DR: "We present Matrix3D, a unified model that performs several photogrammetry subtasks, including:
- Pose estimation
- Depth prediction
- Novel view synthesis
All using the same model."
Contributions (cited from paper):
In summary, our key contribution is the unified diffusion transformer model, which has flexible input/output configurations and enables several tasks in the photogrammetry process.
The proposed method eliminates the need for multiple task-specific models, streamlining the process with a single model.
Extensive quantitative and qualitative experiments demonstrate the state-of-the-art performance of our method compared with existing task-specific approaches.
3R-GS: Best Practice in Optimizing Camera Poses Along with 3DGS
Zhisheng Huang, Peng Wang, Jingdong Zhang, Yuan Liu, Xin Li, Wenping Wang
tl;dr: MLP pose refiner, 3DGS as sampling (see https://t.co/Yl2ppI3Y1R), epipolar contraints
https://t.co/uRio6RlDnU https://t.co/blsYLzfgGr
This is how I generate the scores for ranking: https://t.co/tYKo1J54kd
The agent vs the orchestrator.
The agent is commodity labor. It is increasingly reliable at completing well-defined tasks but (by itself) easily replaceable.
The "orchestrator" sits a layer above the agent. The orchestrator absorbs your company's process knowledge (how do we handle an anticipated scenario?) The orchestrator handles access control, observability, and other things that resemble HR functions back when the resources were human.
The labs are all going to build agents. Some of these agents will be vertical-oriented, e.g. for software. Some will be general purpose, taking form in the apps we use, armed with email addresses and Slack (or whatever comes after Slack) accounts.
But it will be ill-advised for companies to offload administrative/orchestration functions to a source of agent labor. Many companies (not all) will resist lock-in from agent companies and favor systems of work that embrace agent-neutrality.
All of which is to say, I am generally more interested in investing in companies that operate at the orchestration/administrative layer than the agent layer.
you can now get the research app for macOS, windows and linux right from the homepage ☁️ https://t.co/IFLUd46Ktx https://t.co/GysbBBQTmn
The most underrated AI superpower isn't fancy prompting—it's writing down your workflow.
1. Record your process
2. Define what good looks like
3. Transform one-off AI experiments into systems that can be automated and shared with your team https://t.co/6G0phddA1C
We wrote down everything we've learned building voice AI agents over the past two years.
Core technology choices, minimizing latency, managing multimodal context, interruption handling, turn detection, evals, state machines, guardrails, memory, async and realtime function calling, ...
Plus diagrams, charts, and code samples.
The text is open source, so feel free to submit PRs.
seeing discussions on generalized memory…today would be a great day to finish this blog post haha…but I’m doing other stuff so I’ll just tweet w an excerpt
Dot has 4 types of memory & 4 ways it fetches across them. We also made it “dream” to compress & retire memories https://t.co/w0OKaUyNyl
We're excited to announce Cloudflare Realtime and RealtimeKit, a complete toolkit for shipping real-time audio and video apps in days with SDKs for Kotlin, React Native, Swift, JavaScript, and Flutter. https://t.co/y5lSmYFgW5 #DeveloperWeek
.@cloudflare developer week, day 2: a cloudflare worker can contain multitudes...
→ you can use a cloudflare worker to host static sites, SPAs or SSG apps
→ supported frameworks: react router 7, astro, hono, vue, nuxt & more
... too big for one tweet, continued in thread 🧵