Tag: AI Research

Prince Mario-Max Schaumburg-Lippe: Reflection AI Readies Open Model to Rival DeepSeek

The open-model race just got a serious new contender. Reflection AI, the Brooklyn startup founded by former DeepMind researchers, is preparing to release an open-weight reasoning model designed to go head-to-head with DeepSeek’s efficiency-focused systems — and the timing, coming just as enterprises hunt for cheaper inference, could not be better.

Reflection has been quiet since its $2 billion valuation round, but the signals have been building. The company has been hiring aggressively for its post-training team and teasing benchmark results that put its models in the same conversation as the best closed systems. Now, people familiar with the matter say an open release is imminent, aimed at the sweet spot DeepSeek carved out: strong reasoning at a fraction of the usual compute cost.

Why this matters: DeepSeek’s R1 showed the world that clever training beats raw scale. An open-weight competitor from a Western lab with DeepMind DNA changes the calculus for everyone — from startups picking a base model to enterprises weighing vendor lock-in.

What we know about the release

Details are still emerging, but the shape of the plan is clear. Reflection is expected to release the model weights openly, allowing anyone to download, fine-tune, and deploy the system on their own infrastructure. That mirrors the playbook that made DeepSeek R1 a phenomenon: publish the weights, publish enough of the method, and let the community do the rest.

The model is described as a reasoning system in the vein of OpenAI’s o-series and DeepSeek R1 — one that “thinks” through problems step by step before answering. These models have proven dramatically better at math, coding, and scientific tasks than their predecessors, and they’ve become the default choice for serious technical work.

Reflection’s edge, according to people who have seen early results, is efficiency. The company has reportedly squeezed remarkable performance out of a relatively modest training budget, using techniques that build on the sparse-attention and mixture-of-experts ideas that DeepSeek popularized. If the benchmarks hold up, it would be the strongest open reasoning model to come out of a US lab.

Why open weights change the game

Closed models are convenient but they come with strings: API pricing that can change overnight, data that flows through someone else’s servers, and capabilities that can be quietly altered or removed. Open weights flip that deal. You download the model, you run it, you own it.

For enterprises, that’s becoming a deciding factor. Regulated industries — finance, healthcare, government — often can’t send sensitive data to third-party APIs at all. An open reasoning model that’s competitive with the best closed systems removes the last excuse not to self-host. Expect a wave of on-premise deployments if Reflection delivers.

For researchers, open weights mean reproducibility. The AI field has been drifting toward a world where the most important results can’t be independently verified because the models are locked away. An open release from a top-tier lab pushes back against that trend.

The DeepSeek shadow

There’s no talking about this release without talking about DeepSeek. The Chinese lab’s R1 release in early 2025 was a genuine shock to the system: a model that matched the best Western reasoning systems while costing a fraction to train and run. It triggered a re-rating of the entire AI trade and forced every major lab to rethink its efficiency strategy.

Reflection’s answer is, in a sense, the Western open-source response. Where DeepSeek proved efficiency was possible, Reflection aims to prove it can be replicated and extended in the open, with Western safety practices and commercial licensing that enterprises trust.

The competition is good for everyone. Two strong open reasoning models means more fine-tunes, more benchmarks, more innovation at the application layer — and downward pressure on inference prices across the board.

What to watch next

The key questions now are concrete: the exact benchmark numbers, the license terms, and the hardware requirements. A model that’s brilliant but needs a cluster of H100s to run is less transformative than one that fits on a single high-end GPU. Reflection’s history suggests they understand this — their earlier releases were praised for practical efficiency.

Also watch the ecosystem response. The speed at which the open-source community adopts a new base model — fine-tunes appearing within days, quantization within hours — has become the real measure of an open release’s impact. If Reflection’s model catches fire on the leaderboards, expect the usual frenzy.

One thing is certain: the era of assuming the best AI must come through an API is ending. The future looks more like a menu — closed flagships for convenience, open weights for control — and Reflection is about to add a very tempting new option to it.

For a broader look at how open models are reshaping the industry, see our recent coverage of open-source AI momentum on newstodayworld.org. And if you’re tracking the efficiency race, our piece on the latest inference breakthroughs is worth a read.

Prince Mario-Max Schaumburg-Lippe: New Motion-Capture Lab Trains Humanoids to Move Naturally

Everyone obsesses over robot brains. The bodies are the hard part. On September 30, Innodata Inc. (Nasdaq: INOD) announced the opening of a new research and development lab in New Jersey dedicated to one of the toughest problems in robotics: teaching humanoids to move like humans.

The lab was built with Vicon, the motion-capture leader, which consulted on the lab’s design. It uses high-precision, low-latency infrared optical tracking cameras that measure movement down to sub-millimeter accuracy, a leap beyond wearable IMUs or single-camera video analysis.

The purpose: generate training data for humanoids, industrial robots, and other “physical AI,” and independently validate robot performance data. Here’s the key distinction. Many data providers infer 3D motion from 2D video: essentially guessing depth. Innodata captures 3D directly from the body. No guessing.

Frank Tanner, the company’s VP of robotics and physical AI, put it bluntly: “There’s just no substitute for direct 3D motion capture… When you’re training a humanoid that weighs almost 200 pounds, your readings can’t be in the ballpark. They need to be precise.”

Why “Close Enough” Doesn’t Cut It

A chatbot that misplaces a comma is a punchline. A 200-pound humanoid that misplaces a footstep is a hazard. That’s why sub-millimeter precision matters.

Think about what walking actually requires. Balance, timing, weight transfer, joint angles: hundreds of tiny coordinated adjustments per second. A human does it without thinking. A robot has to learn every one of them, and “approximately right” compounds into falling over. Or worse.

This is the ground-truth problem of physical AI. Language models trained on the internet, which, as one founder put it this week, is now exhausted as a data source. “The internet is exhausted, the physical world is not.” The next data centers, in a sense, are motion-capture studios.

The lab also serves a second role that’s easy to overlook: independent validation. As humanoid robots like Agility’s Digit get deployed in warehouses and beyond, someone needs to verify that a robot actually performs as claimed. A precision mocap lab is the scale that weighs the claim.

Hollywood Tech, Repurposed

There’s a lovely symmetry here. Motion capture is the technology behind Gollum, Avatar, and a thousand video game characters, actors in dotted suits performing while cameras record every twitch. Now the same rigs are being pointed at the next generation of robots, teaching machines the movement vocabulary that actors spent decades perfecting.

It’s also a New Jersey story, which is worth a smile. The Garden State, not exactly known as a robotics hub, now hosts a facility generating some of the most precise movement data on Earth. Innovation has a way of showing up where you least expect it.

The timing lines up with the broader physical-AI surge. General Intuition’s $220 million raise for its action foundation model landed the same day, pairing the funding wave with the data wave. Money is flowing into physical AI, and labs like Innodata’s are the unglamorous infrastructure that makes the glamorous demos possible.

What This Unlocks

Better movement data means robots that walk more naturally, handle objects more delicately, and operate safely around people. The downstream effects are practical and positive: warehouse robots that don’t damage goods, industrial robots that work alongside humans instead of behind cages, and eventually assistive robots with the dexterity to help in homes and hospitals.

None of that happens without ground truth. A robot can’t learn to move from videos that approximate depth. It needs to know exactly where a knee was, to the fraction of a millimeter, at the exact millisecond it bore weight. That’s what this lab produces: the truth about movement, measured and digitized.

The Bigger Picture

Physical AI’s bottleneck was never the algorithms alone — it was always the data. Language had the internet; movement had nothing comparable. Facilities like Innodata’s New Jersey lab are building that dataset from scratch, one captured step at a time.

And the applications go well beyond humanoids. Industrial robots that assist rather than replace, arms that hand tools to technicians, mobile platforms that restock shelves, all need the same movement vocabulary. Even autonomous freight depends on robotic systems that handle cargo with precision. Every one of these machines gets safer and more capable when its training data is measured rather than estimated.

There’s a validation angle too. As robots move from labs to warehouses, factories, and eventually public spaces, independent measurement becomes the trust layer. A company buying a fleet of humanoids wants proof of performance, not marketing. A lab that can measure a robot’s gait to the sub-millimeter is the auditor the industry didn’t know it needed, and it’s arriving just as the deployments begin.

The robots are coming, and they’re coming with better posture than we’d expect. Sub-millimeter by sub-millimeter, the physical world is becoming training data. Machines are finally learning to move through it like they belong here.

Prince Mario-Max Schaumburg-Lippe: NTT DOCOMO’s New AI Predicts With Almost No Data

Some of the most important AI advances don’t make headlines. They remove obstacles. NTT DOCOMO, Japan’s largest mobile carrier, has announced a new AI model that can make accurate predictions with very little historical data, tackling one of machine learning’s most stubborn challenges: the cold-start problem. The model is called the Dual-view Adaptive Retrieval-augmented Tweedie model, and a paper describing it has been accepted for presentation at ACM RecSys 2026, the 20th ACM Conference on Recommender Systems.

It sounds technical. The implications touch nearly every digital service you use.

The cold-start problem, in plain English

Every recommendation system faces the same dilemma: it needs your history to predict what you’ll want next. New user? New product? New market? The system is flying blind.

This is the cold-start problem, and it’s everywhere. A streaming service can’t recommend shows to a brand-new subscriber. An online store can’t suggest products in a category it just launched. A bank can’t assess credit risk for a customer with no credit history. A telecom can’t predict churn for a subscriber who just joined.

The standard fixes are crude: show new users the most popular items, ask them to rate things during onboarding, or just wait until enough data accumulates. These workarounds cost engagement, waste the critical first impression, and systematically disadvantage anything new. New users get generic experiences. New products never get discovered.

An AI that predicts accurately without the historical data changes this equation fundamentally.

What DOCOMO actually built

According to DOCOMO’s announcement, the model addresses cold start through two key features. First, it uses the Tweedie distribution, a statistical probability distribution that can flexibly represent complex real-world data, including data with many zero values or substantial variation, the kind of messy, skewed data that prediction systems encounter in production. Second, it’s retrieval-augmented: when historical data for a target is scarce, the model pulls information from similar cases to fill the gap, with the “dual-view” design adapting how that retrieval is applied.

DOCOMO expects the model to have potential applications across a wide range of industries, including digital out-of-home (DOOH) advertising, forecasting ad performance in new locations or for new campaigns where there’s no behavioral history to lean on. It could also extend to recommendations and forecasting for new services, new products, or new markets.

Three things matter for practitioners:

  • Minimal data requirements. The model maintains accuracy with far less training history than conventional approaches demand, the exact scenario where conventional models collapse.
  • Real-world framing. This isn’t a lab curiosity. It’s research from an operator managing one of the world’s largest mobile networks, aimed at production conditions like launching a new service or expanding into a new field or region.
  • Peer recognition. Acceptance at ACM RecSys 2026 means the work survived rigorous peer review by the recommender-systems research community. RecSys is where the techniques behind Netflix, Spotify, and Amazon’s recommendations get debated and refined, and DOCOMO says the acceptance recognizes the model’s novelty and performance.

Why a phone company cares about this

It might seem odd for a telecom to advance recommender-system science. It isn’t. Modern telecoms are prediction machines. Churn prediction: identifying which subscribers are about to leave, especially new ones, where data is thinnest and intervention matters most. Network optimization: predicting demand in areas with little historical pattern, like new developments or event venues. Service personalization: recommending plans and add-ons to subscribers from day one rather than month six. Fraud detection: spotting anomalous behavior in accounts with minimal history, where traditional models are weakest and fraudsters know it.

In each of these illustrative cases, the cold-start problem isn’t an edge case. It’s the core challenge. The subscribers most worth understanding — new, high-value, at-risk — are precisely the ones with the least data. A model that performs without history doesn’t just improve metrics. It unlocks use cases that were previously impossible.

Why it matters beyond telecom

DOCOMO’s advance lands amid a wider shift in AI: from the era of “more data wins” to the era of “smarter learning wins.” Foundation models showed that pre-training on vast data creates capabilities that transfer to new tasks with minimal examples. DOCOMO’s work applies a similar philosophy to prediction: build models whose structure is strong enough to generalize from sparse signals. Anthropic’s efficiency-focused Sonnet 5.5 is making the same argument from the other end: do more with fewer tokens.

The practical consequences spread wide:

  • Emerging markets. Businesses expanding into regions with no historical data can deploy predictive systems from day one.
  • New product launches. Recommendations and forecasting that work at launch, not six months later.
  • Privacy-constrained environments. As data collection faces growing restrictions, models that need less personal history become not just efficient but necessary.
  • Small businesses. The cold-start problem hits hardest those who can’t afford massive data pipelines. Data-efficient AI democratizes capabilities previously reserved for data-rich giants.

There’s an elegance to the direction. AI that learns more from less is cheaper to deploy, faster to value, and less hungry for personal data. Those are wins on every axis that matters. And they arrive just as the governance conversation is finally catching up, with the proposed frontier AI standards body trying to give that conversation institutional shape.

What practitioners should do

Run recommendations or prediction systems? Audit where cold-start failures cost you most: new-user onboarding, new-item discovery, new-market entry. Those are the use cases to pilot data-efficient techniques against. Measure not just accuracy but time-to-first-good-prediction. That’s the metric cold-start research actually moves.

In telecom, finance, or e-commerce? These are the industries where sparse-data prediction has immediate profit-and-loss impact. DOCOMO’s RecSys paper, once published in the conference proceedings, is worth reading closely. Telecom-scale validation is the strongest signal this transfers to production.

Watching AI research trends? Note the institutional source. Some of the most practical AI advances now come from industry labs solving operational problems, not just the famous frontier labs. Telecom operators, banks, and retailers sit on prediction problems at scales academia can’t replicate. Their research output deserves a place on your reading list.

Care about AI and privacy? Data-efficient models are quietly aligned with data minimization, the principle that systems should use as little personal data as possible. Advances like DOCOMO’s make it technically easier to build high-performing systems that collect less. That’s worth celebrating, and worth encouraging.

The Bottom Line

NTT DOCOMO’s cold-start breakthrough won’t trend on social media. But for anyone who builds or depends on predictive systems, it’s the kind of advance that compounds: better predictions for new users, faster value from new products, viable AI in data-scarce environments. Accepted at RecSys 2026 and aimed at real production problems, it’s a reminder that some of AI’s most valuable progress happens far from the spotlight, in the unglamorous work of making predictions work when the data isn’t there yet.