The Lambda Deep Learning Blog

From tokens to concepts: how particle models perceive the world

amirali.zadeh@lambda.ai (Amir Zadeh) — Mon, 03 Aug 2026 12:30:30 GMT

Patch-based vision models exhibit a related class of failure modes: fixed patches may split a single object across multiple tokens or place parts of several objects within one token. This, in turn, makes it difficult for the model to infer object boundaries and correctly associate features across patches—an instance of the broader visual binding problem in computer vision.

Prompt injection doesn't care what your agent does for a living

Devina Jain — Fri, 31 Jul 2026 18:34:23 GMT

What 1,433 winning attacks looked like when we clustered them

Most teams test agent security one domain at a time. Is the customer-support bot safe? The code assistant? The expense approver? Each gets its own red-team pass, its own scenario list, its own sense of "we checked."

Keeping 100k battles of untrusted agent code in their lane

David Hartmann — Thu, 30 Jul 2026 12:26:21 GMT

In March 2026, Lambda ran AgentBeats, an AI agent security competition in which teams submit two kinds of agents: an attacker that tries to manipulate a target LLM into doing something harmful, and a defender that tries to stay helpful while refusing the trap (check the final leaderboard here). Our platform pairs them, runs the battle, and scores the outcome. The agents are graded on how effectively they subvert the system, which means the platform's job is to stay correct and on schedule while the code it hosts is trying to break things.

In high-frequency trading data, noise isn't the problem. Assumptions are.

Jessica Nicholson — Fri, 24 Jul 2026 12:25:43 GMT

Lambda recently teamed with Hudson River Trading (HRT), one of the most respected quantitative trading firms in the world, to power their trading research and development on Lambda Cloud. It's a deal that reflects something we've been seeing more broadly: access to compute is a necessary condition for frontier quantitative research, but it's not sufficient. The firms doing the most ambitious work are running into a separate, harder problem in how they prepare and understand their data.

Your coding harness shouldn't be a black box

Zach Mueller — Tue, 21 Jul 2026 19:01:02 GMT

New models ship every day, both open and closed. The harness you run them through decides how much of that capability you actually get. Depending on the harness, you can get wildly different results from the same model. A smaller model with the right tools can outperform a larger one, and a harness tuned for the right model can blast away its competition.

One photo in. A full 3D scene out.

Jianwen Xie — Mon, 20 Jul 2026 12:21:17 GMT

Point a camera at a room. Get back a complete, editable 3D model of everything in it: the sofa, the table, the chairs, each one a clean mesh you can rotate, move, and drop into a game engine or a design tool. That is PixARMesh, a single-image 3D scene reconstruction method from UC San Diego and Lambda, accepted at CVPR 2026.

Why your Kubernetes scheduler can't handle AI workloads

Cody Brownstein — Thu, 16 Jul 2026 12:35:42 GMT

Imagine this scenario: You have a distributed training job with 16 worker pods, each requesting 1 GPU. 4 GPUs are currently available. The default Kubernetes scheduler (kube-scheduler) may schedule those 4 pods while the remaining 12 stay pending.

Meanwhile, those 4 GPUs are reserved by pods that cannot make progress until the full distributed job is ready. No one else on the cluster can use them. Your job isn't actually training. The other cluster users' jobs can't use those reserved GPUs either. The cluster is busy doing absolutely nothing.

This is a classic partial-scheduling deadlock. By default, kube-scheduler doesn't gang-schedule; the feature exists in alpha but ships disabled. It schedules pods one at a time, meaning it has no concept of "all pods in this job must start together, or none of them start."

The second problem is that kube-scheduler lacks multi-node fabric topology awareness. While it can track basic on-node constraints such as available CPU or memory capacity, it doesn’t account for which separate nodes share the same NVIDIA Quantum, or how high-bandwidth interconnects are routed across the cluster network. It places pods wherever capacity exists. For distributed training, this introduces severe communication latency that throttles performance at scale.

For teams running small jobs on a few GPUs, this inefficiency is tolerable. For AI/ML organizations running distributed training across dozens or hundreds of GPUs, it breaks the entire workflow.

GLM 5.2: a new rise of open-weight agentic models

Zach Mueller — Thu, 09 Jul 2026 12:25:13 GMT

On June 16th, Z.ai released GLM 5.2, its latest flagship model. At the time of announcement, it advertised scores at or near Anthropic and OpenAI's models, and far ahead of GLM 5.1. In the world of usable, deployable, and reliable AI models, however, the benchmarks matter, but don’t paint the picture of how capable the model is in the real world. What happened next has turned the field on its head.

Many call it the "DeepSeek moment for agents."

GLM itself is the exact same architecture at 744 billion parameters. So large that most individuals can’t run it on their own hardware and must rely on cloud compute just to access it. With large models such as these, the usual pattern is that for a few days everyone gets excited, a few individuals run it on-premises and show how slow (or fast) it can run quantized, and then most people fall back into the cloud with models like Anthropic's Claude Opus and OpenAI’s GPT. More reputable sources reported not only that this model is the real deal, but also that experienced labs and industry leaders were replacing much of their workloads with GLM (and, a few weeks later, keeping it there after extensive testing).

Lambda’s keynote at the ALVR workshop co-located with ACL 2026

amirali.zadeh@lambda.ai (Amir Zadeh) — Wed, 01 Jul 2026 21:03:32 GMT

On July 3, Lambda's research team delivers a keynote at the Advances in Language and Vision Research (ALVR) workshop, co-located with ACL 2026 in San Diego.

What happens when Claude Code gets an experiment tracker

David Hartmann — Thu, 25 Jun 2026 14:58:43 GMT

At CVPR 2026, Lambda ran a live demo for two and a half days: Claude Code teaching Google's Gemma 4 to play a Tetris-like game. Claude Code started with a Gemma 4 model that couldn't play at all. It pressed “down” and lost in seconds. By the end of the demo, the same agent had iterated through hundreds of experiments, trying various board representations, image inputs, prompts, inference settings, and vLLM launch parameters. Slowly but surely, Claude taught Gemma 4 how to play.

No human tuned the parameters. No human chose the prompts. Claude Code ran the experiments, tracked what worked, and built on its own results.

It ran on GPUs that would’ve have been underutilized otherwise: 468 experiments for zero added compute cost.

The tool that made this possible is the_lab.api.