AWS lets you fine-tune Nemotron 3 without touching a GPU cluster

Good morning 👋 Serverless fine-tuning just picked up the model everyone's actually asking about. AWS still won't run the cost math for you, though.

In today's issue:

• 🔭 AWS fine-tunes Nemotron 3, no GPU cluster required
• 🧠 AWS's inference-engineering trifecta, plus OpenAI's telecom playbook
• 🔬 Agent case management, a benchmark that watches instead of grades, and reasoning by video frame
• 🏛️ Aircraft tariffs teased not fired, and the EU prices its own Fund
• 🛠️ Fine-tune Nemotron 3 without standing up infrastructure, plus a CLI fork worth a look
• 🔗 A GPU-hosting pivot, a reverse split, and a $285-in-the-bank filing

🔭 THE ONE THING

🛠️ AWS lets you fine-tune Nemotron 3 without touching a GPU cluster

AWS just expanded SageMaker AI's serverless customization to cover both NVIDIA Nemotron 3 models, Nano (30B total, ~3B active) and Super (120B total, ~12B active), and this time it's not just supervised fine-tuning. RLVR and RLAIF are on the menu too, per the blog post, a real step up from the Nano-only, SFT-only version AWS shipped back in June. The pitch: submit a training job, get a checkpoint back, no cluster to provision or babysit. That's genuinely useful if fine-tuning has been off the table because standing up multi-GPU training infra felt like its own project. Whether serverless beats renting the boxes yourself once your jobs run daily instead of occasionally is the math AWS's post conveniently doesn't do for you.

🧠 MODELS & RELEASES

• ⚙️ AWS dropped three inference-engineering posts on its ML blog in one day: four ways to deploy Unsloth-quantized models on SageMaker (EC2 and managed endpoints both covered), a KTern.AI writeup on running agentic SAP workflows on Bedrock AgentCore with persistent context and secure tool access, and a guide to disaggregated prefill/decode on SageMaker HyperPod using vLLM. The last one's the read if you're running inference at any real scale. Splitting prefill and decode onto separate GPU pools is how you stop paying for compute your workload isn't actually using. Unsloth on SageMaker · KTern.AI on AgentCore · disaggregated prefill/decode
• 📞 OpenAI published its Deutsche Telekom case study: customer service, employee workflows, network operations, even voice, rebuilt around its models. Treat it less as a customer testimonial and more as OpenAI's enterprise telecom playbook, spelled out for whoever's next. Deutsche Telekom

🔬 RESEARCH HIGHLIGHTS

• Quick Automate's case management AWS wrapped agent runs in first-class "cases" that move through a real lifecycle (Ready, In Progress, Successful, Failed, Pending Resolution), with a creator/processor split so one automation ingests work and many process it in parallel. That's less a research finding than a naming exercise, the one enterprises actually need before they'll trust a fleet of agents with production work and a paper trail. Post
• UniClawBench checks whether agents can do things, not recite them. 400 bilingual tasks run live in Docker containers, scored step by step across five capabilities (skill use, exploration, long-context reasoning, multimodal understanding, cross-platform coordination), instead of graded off a static answer key an agent can game. Grading criteria stay hidden during the run. Closer to how you'd actually judge an assistant. Paper
• OpenCoF swaps chain-of-thought text for chain-of-frame video. The model reasons by generating a sequence of images instead of writing out steps, trained on a newly built dataset for the task. The resulting model, Wan-CoF, beat its baseline across four reasoning benchmarks. Early-stage work, but it's a real candidate for tasks where the reasoning is spatial, not linguistic, think physical setups a paragraph struggles to describe. Paper

🏛️ POLICY & REGULATION

• The White House teased Section 232 tariffs on imported commercial aircraft, jet engines, and parts, then held off. Commerce recommended against immediate tariffs, so the proclamation just directs Commerce and USTR to go negotiate with trading partners, with a 180-day progress check written in. No tariff exists yet. File it as a warning shot, not an action item. Proclamation
• EU Commission put real numbers behind its own Competitiveness Fund pitch. JRC and DG Connect modeling says every euro spent through the Fund's Digital Leadership window returns €1.84 to €2.21 of EU GDP over ten years, roughly €13-16 billion a year by 2037. A quarter of each country's gain comes from spillover off someone else's spending. That's the real argument for running this centrally instead of as 27 separate national programs. Study

🛠️ TRY THIS

Fine-tune a Nemotron 3 model without standing up infrastructure

1. Pick the smallest Nemotron 3 variant that clears your accuracy bar. Bigger isn't safer, it's just slower and pricier to serve.

2. Assemble a small, high-signal fine-tuning set. A few hundred clean prompt/completion pairs beat a sloppy 50k. JSONL, uploaded to S3.

3. Launch the job through SageMaker AI's serverless customization path. No endpoint to provision, no cluster to babysit. You pay for training compute and it tears itself down when the job ends.

4. Evaluate against a held-out slice before you deploy. A fine-tune that nails your support tickets and tanks on everything else isn't done yet.

Prompt: Given this fine-tuning dataset of {N} examples, flag duplicate or near-duplicate prompts, inconsistent labels, and formatting errors that would hurt a Nemotron 3 fine-tune. Output a cleaned JSONL plus a list of what you removed and why.

Worth a look

• ai-job-search is the outlier here: a Claude Code fork that scrapes postings, drafts LaTeX cover letters, and preps interviews end to end. 20.7k stars. Less useful for the job hunt, more as a clean template for a drafter-reviewer agent pipeline. github.com/MadsLorentzen/ai-job-search

🔗 QUICK LINKS

• Azio AI Holdings (the EV maker formerly known as Envirotech Vehicles, ticker EVTV) landed a $27.9M GPU hosting deal with Power Champion Investment, 3.1 MW now with room to scale to 12, its first real revenue proof for the pivot from vehicles to AI data centers. Filing
• LM Funding America (LMFA), the bitcoin miner turned AI compute host, is running a 1-for-25 reverse split to stay Nasdaq-listed while it leans harder into its 26 MW Oklahoma and Mississippi buildout. Filing
• Bravalo Corp filed its fourth S-1 amendment for "HeadLab," an AI headline generator: $285 in the bank, zero revenue, raising $125,000 at two and a half cents a share. Filing

See you tomorrow.

Pradeep Perugu