OpenAI Gives Agents a Wallet, Nvidia Faces a CEO Crisis
7 min read · 16 sources
- OpenAI introduced ChatGPT Wallet, letting agents make purchases with a security check on every transaction and without exposing card details.
- Devin Apps went live, hosting private internal web apps built from text prompts on Devin's cloud with billing based on awake time.
- Thinking Machines published details on end-to-end 4-bit training, with NVFP4 running 6x faster than BF16 on NVIDIA GB300.
- Quark, a 165M-parameter model, was trained end-to-end for about $4,200 of compute and released under Apache-2.0.
- A petition to reinstate ousted Nvidia CEO Jensen Huang gathered over 115,000 signatures, including from Elon Musk.
OpenAI wants your AI agents to be able to buy things, but it doesn’t want them to have your credit card number. That’s the pitch behind ChatGPT Wallet, which landed today and gives agents a spending mechanism that requires a security check on every single transaction. It’s the kind of infrastructure that turns agents from chat companions into something that can actually close a deal.
And while OpenAI was handing out wallets, Nvidia was dealing with the fallout of losing its CEO. A petition to reinstate Jensen Huang has crossed 115,000 signatures, with Elon Musk and other tech executives publicly backing it. The company’s strategic direction is suddenly in question, and that matters for every engineer who’s been building on the assumption that Nvidia’s roadmap was set in stone.
NVFP4 runs up to 6x faster than BF16 on GB300, but the quantization error it introduces needs careful management across training stages.
ChatGPT Wallet: Agents Get a Card, You Keep the Number
ChatGPT Wallet is OpenAI’s answer to the question of how an agent pays for things without you handing over your card details. The wallet lets agents make purchases, but every transaction requires a security check. Your actual card number never touches the agent’s hands.
This is a meaningful shift. Up until now, agentic commerce has been stuck in a loop: either you pre-fund some sandboxed account, or you let the agent loose with your real payment details and hope it doesn’t go rogue. The wallet sits in between, giving agents spending power while keeping the underlying payment rail private. For engineers, the interesting bit is the security check on every transaction - that’s a human-in-the-loop guardrail that will shape how agent workflows get built.
Devin Apps: Internal Tools From a Text Prompt
Devin Apps are live. These are private web applications that Devin builds and hosts on its cloud computer, generated from a text description or prompt template. They get stable URLs, require Devin sign-in, and can be shared within an organization.
The architecture is worth noting. Each app runs on a single machine using memory snapshots, sleeping when idle and waking on access. Billing is based on awake time, not wall-clock time. The apps support any HTTP-serving stack, databases like SQLite and Postgres, and the Devin API. That means an engineer can spin up an internal dashboard that has access to existing MCPs, integrations, and codebases without standing up a whole deployment pipeline.
The catch: this is not for high-traffic production services. A single machine that sleeps when idle is fine for an approval queue or a monitoring dashboard, not for something serving millions of requests. But for internal tooling, it collapses the time from “I need a dashboard” to “here’s a URL” down to a single prompt.
End-to-End 4-Bit Training: The Speed/Error Tradeoff
Thinking Machines published a deep dive into training language models end-to-end at 4-bit precision on NVIDIA’s Grace Blackwell Ultra (GB300). The piece covers three numerical formats: BF16, MXFP8, and NVFP4.
NVFP4 is the headline. It uses 4.5 bits per value and runs up to 6x faster than BF16 - 15 PFLOPS versus 2.5 PFLOPS on GB300. But it introduces significantly more quantization error. The trick is that this tradeoff isn’t uniform across the training pipeline. Pretraining forward/backward passes, SFT, and RL each have different priorities when it comes to minimizing error, bias, and memory footprint.
For engineers actually training models, this is the difference between a 6x speedup that breaks your model and a 6x speedup that works. The guidance here is to be surgical about where you apply low precision. Apply it in the wrong stage and you’ll watch your eval scores crater; apply it where error is tolerable and you’re saving real money on compute.
Quark: A Full Training Stack for $4,200
Quark is a 165M-parameter language model (140M without embeddings) with 128K context, trained end-to-end from raw data through pretraining, long-context training, RL, evaluation, and release. The entire thing cost about $4,200 of compute, roughly 1,300 H100-hours.
The architecture is 16 layers with 12 local layers attending to the last 256 tokens and 4 global layers, plus multi-head gated value residuals. One version, Quark-150M-128K-Zero, learned agentic tasks from RL alone - no SFT, no distillation. All four checkpoints are released under Apache-2.0.
This matters because it validates the whole pipeline before you spend serious money. The team explored every part of the stack - data, architecture, pretraining at scale, context extension, RL, evaluation - on a budget that’s less than a single large training run. If you’re planning to train something bigger, this is the proof-of-concept playbook.
On the Brutality of Amdahl's Law
Source: bosoncutter.substack.com ↗
A piece on Amdahl’s Law argues that the reason AI hasn’t caused rapid economic transformation is that intelligence isn’t the bottleneck. In experimental science, and in most real-world workflows, the surrounding systems are just too slow.
The author notes that Claude Code + Opus 4.5 in December 2025 made agents feel genuinely useful, but predictions of rapid white-collar unemployment were wrong. The world has seen only 2% GDP per capita growth for nearly 150 years, despite electricity, communication, computers, transportation, and agriculture all going “quasi-infinite.” Intelligence is just another curve, and it’s hitting the same wall of everything else being slow.
This is a useful corrective for anyone who’s been assuming that a model that beats humans on benchmarks will immediately translate to economic output. The model isn’t the bottleneck. The rest of the organization is.
Qwen4-Max: Your New Teammate Has Arrived
Qwen4-Max is positioning itself as a generalist that can handle 3D modeling, game development, precision manufacturing, finance, law, healthcare, and frontier scientific discovery. It can use Blender for geometry modeling, scene assembly, material setup, lighting, and rendering. It also works with leading game engines, handling scene construction, gameplay logic, external asset integration, material setup, HUD design, and visual effects - and can deliver a playable game.
The open source angle matters: smaller models in the family are coming. If the capability gap between the flagship and the open models is small enough, this becomes a real option for teams that want on-prem Qwen without the API cost.
Embla 1: A Reasoning Model That Runs on Your Phone
Embla 1 is a reasoning model designed to run on phones - private, offline, and free per request. At 4 bits, it averages 57.4 on BFCL, IFBench, AA-Omniscience, GPQA, and MATH-500, with a 1.77-second end-to-end response on an iPhone 18 Pro.
The architecture is 32 layers of width 2048 with 8 of 96 routed experts, using Sevren Sparse Attention - ReLU² instead of softmax, with roughly 95% zero weights - plus shared K/V caches and an Engram hash table. Training included 1.4T token pretraining, 20B token context extension to 256K, 123B token SFT with QAT, and RL stages including GRPO and distillation.
For engineers, this is the first credible sign that on-device reasoning is viable. If a model of this quality runs offline on a phone, a lot of the privacy and latency arguments for cloud-only AI start to weaken.
BirdDreams: A Driving World Model From NVIDIA
BirdDreams is an NVIDIA driving world model that predicts driving video and a bird’s-eye view from a single camera image, with optional ego trajectory control in meters. It generates RGB and BEV videos without HD-map or surrounding-vehicle-state inputs, and traffic reacts to ego maneuvers.
In human evaluations - 25 participants, 375 pairwise judgments - BirdDreams showed more realistic vehicle behavior than an OmniDreams-based baseline. For autonomous driving simulation, this enables precise ego control and reactive traffic prediction, which is exactly what you need to test planning algorithms without deploying cars.
Personal Agent Protocol 1.0: Agents Talking to Businesses
Poppy v1.0 is now available. The Personal Agent Protocol enables personal agents to interact with companies through business agents, websites, and APIs. Shopify has enabled Poppy for every merchant, and Meta’s Muse agent now looks for Poppy first on every site. Over 60 new companies joined as design partners, bringing the total past 100, including Salesforce, Delta Air Lines, and Zoom.
The protocol uses existing standards like MCP, OpenAPI, and UCP with OAuth authentication, plus a commerce extension co-authored by Shopify. If this standard takes hold, it reduces the integration friction between personal agents and businesses to a single protocol handshake.
Quick Hits
GPT-6 Astra worked with physicist Nathaniel Craig to develop a candidate lattice formulation of the Standard Model, addressing a longstanding challenge without perturbative approximations.
Anthropic published recommendations for building strong safety evaluations, drawing on psychology’s replication crisis reforms including preregistration and rigorous statistics.
Claude Code got mods - TypeScript plugins installed with /plugin that change behavior, customize UI, and swap features. Opus 5.5 is now default for paid plans, roughly 30% faster and 40% cheaper than Opus 5.
Meta researchers published a study on vision encoders in multimodal models, finding that dedicated capacity matters more than pretraining, and that encoder-free models need ~3x compute to match.
Anthropic’s guide on building always-on agents covers the core loop with five primitives: event inbox, timers, memory, permissions, and outbox. Opus 5.5 costs 78% less per solved task on a long-horizon coding benchmark compared to Opus 5.
You May Also Like
Nvidia's board moves to oust Huang as Gemini Decisions API prices decisions at $0.10
Nvidia's board is reportedly preparing to oust CEO Jensen Huang over allegations involving a woman US intelligence believes was a Chinese asset, sending shares …
Nvidia's Board Hunts for Huang's Replacement as Agents Take Over AWS
Nvidia's board has hired an outside firm to identify CEO candidates as it weighs ousting Jensen Huang following reports of extramarital affairs, including one …
OpenAI arrests insider; Google rolls out Gemini 4 Argon; Anthropic’s Claude gets dynamic workflows
OpenAI busted an employee with a hidden recorder, handed the case to the FBI, and confirmed no model assets were stolen. Google Cloud added Gemini 4 Argon, a …




