Bikes that know the block, agents that remember — the week got practical.
Trueforge trims agent costs 30–75% and Kelex gives them long-term memory; Ground Truth reads neighborhood wealth off how hard you pedal. The Simpsons returns, HoverAir flies, Iceberg ditched JSON blobs for a real Variant type — the tools are finally useful. Why this matters for us: cheap agents and real sensors mean the tools we actually use won't keep getting more expensive. }
NYC, carved out of a model, sits in your phone now
A new paper on arXiv (2608.08654) shows how you can use one AI model to generate a 3D model of another. The technique is simple in principle: feed a prebuilt model into a foundation model and let it spit back a new one. What's interesting is the application — the authors made a model of New York City. You can spin it, zoom in, walk around. It's a digital twin of the boroughs, rendered by AI.
This is the kind of thing that sounds like a demo until someone builds the pipeline and starts doing it at scale. The 3D web is getting crowded with these things. GLB files are the lingua franca, and the file sizes are shrinking as the methods improve. Once the models can generate them on the fly — without a human sculpting — the use cases multiply fast.
Why this matters for us: it's another step toward a world where the neighborhoods we live in — the bodegas, the tenements, the bodegas again — can be captured, shared, and used without needing a team of artists behind them.
Trueforge, open-source agent harness, trims costs 30–75% vs Claude managed agents
TrueForge is Truefoundry's open-source framework for running AI agents. It bundles the tools agents need—browsing, code execution, file writes—into one package. The claim: task completion runs 30 to 75 percent cheaper than Claude's managed agents.
Managed agents are the hand-wringing ones that think before they act. They call a model, read the result, call again. Each extra step costs. TrueForge lets agents work in parallel and keeps the loop tight. That's the savings.
It's open source, so you can drop it into your own stack instead of paying a managed fee. For shops already running models and paying for agent overhead, that's a practical hedge.
Why this matters for us: the folks who build tools for the grind are starting to open the gate—so you can run your agents without a vendor's markup.
Tool use: LLMs learning to reach for things
Tool use is how a language model stops talking in a vacuum and starts doing work. The model sees a list of tools it can call—search, code, calendars, APIs—and decides which one to reach for, sends the right arguments, then reads the result back in.
It's not new. ChatGPT has had it for a while; Anthropic's Claude and open-source models like Qwen and Llama 4 all support it. The trick is the schema: you tell the model what tools exist, what they take, and what they return. The model picks the right one and formats the call.
The analogy is a tía who cooks. When you ask her for dinner, she doesn't just guess. She checks the fridge, calls the abuelita for the recipe, pulls the pan from the cupboard. She uses the tools around her and feeds the result back.
Tool use matters because it turns a chatty model into a worker. Instead of writing a long answer from memory, it reaches for what it needs. You get more accurate answers with fewer hallucinations.
For us: if you're using a model that supports tools, check the settings—turn on search or code execution, and give it the right schemas. It'll do the work instead of pretending to know it.
The hardware is fine. The paperwork isn't.
— theverge.com
#the-hoverair-versa-is-flying-but-not-to-the-us-eaac91Nvidia backs Cloverleaf to build more data centers
Nvidia is putting more capital into data center development through a new partnership with Cloverleaf. The deal lines up with Nvidia's own growth — the company is pulling in more cash from the AI data center boom, and now it's helping build the brick-and-mortar side of that bet.
This is the kind of move that keeps the supply chain moving. More data centers means more GPUs flowing in, more rack space getting filled, more revenue cycling through the ecosystem. Cloverleaf brings the real estate and construction chops; Nvidia brings the cash and the demand signal.
Kelex: agents that remember what they've done
Most agent frameworks treat every run as a fresh chat. The agent forgets yesterday, the user gets re-explained things, and the flags pile up with no way to say which ones still matter. Builders who want an actual long-running agent — one that remembers across months, picks up where it left off, and flags what it can't decide — end up writing the substrate themselves.
Kelex is that substrate, productized. It gives you typed memory, tenants and agents as first-class objects, bounded confidence with progressive flagging, and webhooks for human-in-the-loop steering. You get an agent that carries context forward, not one that starts over on every call. We're using it to run Lara and the BFTS content stack before shipping it.
Why this matters for us: the agents we build need to remember the people they serve, not just process the current request. Kelex gives us that without the hand-waving.
https://brownforces.io/solutions
Letara raises $16M to build hybrid rockets beyond satellites
Tokyo startup Letara just closed ¥2.6 billion — about $16M — to fund its next phase. The company is known for hybrid rocket thrusters that nudge small satellites into the right orbit. Now they're moving the same technology into a wider market for space, defense and security hardware.
The raise isn't a moonshot. It's a measured step up from niche thrusters to broader propulsion systems. That's how the work gets done in this space: start with one part, get it right, then scale the platform.
Why this matters for us: Japanese aerospace is quietly building the parts we all need — and they're doing it without the Silicon Valley fanfare.
Apache Iceberg got a Variant type — no more JSON blobs
Apache Iceberg is a table format that sits on top of Parquet, Avro, and ORC. It tracks what changed, when, and who changed it. The latest release adds a Variant type for storing JSON without having to write it as a string.
Before, you had two choices: keep the JSON as a blob and lose the ability to query inside it, or flatten everything into columns and lose the structure. Variant is the middle path. The data stays JSON, but Iceberg can now read individual fields without deserializing the whole thing.
Why this matters for us: as more tools move from blob storage to lakehouse formats, Variant means the data stays queryable — not locked inside a JSON string — and that matters for anything we're building on top of it.
Why this matters for us: it means the JSON we write today won't become a fossil we have to dig out later.
The Simpsons: Hit & Run is coming back — maybe
Matt Groening let something slip at D23: the cult classic game is coming back in some form. Showrunner Matt Selman said "or not" right after him, which is how these things often go. The 2003 original runs on PS2, Xbox, and GameCube, and you can get it working on modern PCs — it's just not on the new consoles.
It's the kind of news you shrug at and then feel a little pang for. The game is loved. The fans want it. The answer is probably yes, eventually, when the right people get the rights straight.
Samsung is making its first over-ear headphones since 2015
Code in the Galaxy Wearable app points to a new pair of over-ear headphones, dubbed the Galaxy H1. They could arrive in 2015, which was the last time Samsung shipped over-ears — the Level On — before it bought Harman and a bunch of audio brands.
The app shows the usual connection icons for over-ear Bluetooth headphones. The sketch isn't the final look, but it's a real design in the works.
Why this matters for us: if they ship in 2027 they'll sit alongside Apple's AirPods Max and Sony's big cans, giving the community another solid pair to buy without paying for the Apple tax.
Ground Truth: a bike sensor that maps wealth by how hard you pedal
Justin Blinder came up with Ground Truth after getting laid off in 2020. He started delivering food by bike around Brooklyn and noticed something most folks ignore: two neighborhoods right next to each other can feel worlds apart economically and psychologically.
The device is a small sensor you clip onto your bike. It reads the effort it takes to pedal up and down the same hills — the steeper the climb, the heavier the load, the more your legs are working against gravity. Those numbers become a proxy for how taxing a neighborhood is.
The Sony A7R VI finally makes high-res practical
The A7R VI is Sony's latest flagship, and it's a 66.8-megapixel sensor monster. The key difference from the last-gen A7R V is a new fully stacked sensor and a Bionz XR2 processor. Together they let the camera do nearly anything — sports, action, usable 8K video — without the sacrifices old high-res bodies forced on you.
Years ago, stacking megapixels meant slower shooting, more noise, and a camera that specialized at the expense of everything else. The A7R VI is the departure from that. It's fast, it's sharp, and it's no longer a niche tool.
Why this matters for us: the aults are getting better without the old compromises, so the family group photo and the kid's soccer clips finally come out clean.