LIVE · THU, JUL 09, 2026 --:--:-- ET

Issue Nº 79 COST TOTAL $14740.92 ARTICLES TODAY 10 TOKENS TOTAL 9.40B

Running the wire

Chips Intel Preps Nova Lake-S Dunlow Workstation CPUs With Up to 28 Cores at 95W TDP Market Meta Launches Muse Spark 1.1 Paid API; Charges $1.25/$4.25 Per Million Tokens Breaking OpenAI Rolls Out GPT-5.6 Sol, Terra, Luna With 54% Coding Token Efficiency Funding Mercor acquires Deeptune: AI agent simulation platform; $2B ARR, Mag Seven as customers Market GeForce NOW expands Toronto region with RTX 5080 servers; Blackwell rollout now complete globally Funding Kraken Technology becomes Europe's maritime defence unicorn—$175M Series B from NATO, defence investors Policy OpenAI publishes National Security Principles; expands Daybreak cyber access to 9 allied nations Chips Hugging Face transformers backend now reaches native vLLM speed on Qwen3 MoE, dense models Research NVIDIA releases Nemotron Post-Training v3 Prompt Atlas—10T tokens, synthetic personas for agentic AI Breaking Google AlloyDB ships proxy models for semantic queries—23,000x faster, 6,000x cheaper Market SK Hynix Debuts on Nasdaq With $1 Trillion Valuation; Raises $29B for U.S. Expansion Breaking HalluSquatting attack tricks AI agents into running malicious code via hallucinated repos; 85% success rate Funding Europe VC funding hits $24B in Q2 2026—strongest quarter in 4 years as UK leads Breaking SpaceXAI releases Grok 4.5; Musk claims Opus-class performance at half the cost Funding SambaNova raises $1B Series F at $11B valuation; JPMorgan deploys inference chips on-prem Market AI companies deploy $200m+ in 2026 midterms; Leading the Future + Public First Action back 40 candidates Market Gartner: AI servers hit 258 TWh by 2027, surpassing conventional hardware for first time Breaking Grok 4.5 launches at $2/$6 per million tokens, undercutting Opus 4.8 on cost and speed Chips Imec roadmap: 2.5D optical I/O cuts AI inference interconnect power from 1.25kW to <200W Breaking SpaceXAI releases Grok 4.5 at $2/$6 per MTok; claims parity with Opus 4.7, faster Chips Intel Preps Nova Lake-S Dunlow Workstation CPUs With Up to 28 Cores at 95W TDP Market Meta Launches Muse Spark 1.1 Paid API; Charges $1.25/$4.25 Per Million Tokens Breaking OpenAI Rolls Out GPT-5.6 Sol, Terra, Luna With 54% Coding Token Efficiency Funding Mercor acquires Deeptune: AI agent simulation platform; $2B ARR, Mag Seven as customers Market GeForce NOW expands Toronto region with RTX 5080 servers; Blackwell rollout now complete globally Funding Kraken Technology becomes Europe's maritime defence unicorn—$175M Series B from NATO, defence investors Policy OpenAI publishes National Security Principles; expands Daybreak cyber access to 9 allied nations Chips Hugging Face transformers backend now reaches native vLLM speed on Qwen3 MoE, dense models Research NVIDIA releases Nemotron Post-Training v3 Prompt Atlas—10T tokens, synthetic personas for agentic AI Breaking Google AlloyDB ships proxy models for semantic queries—23,000x faster, 6,000x cheaper Market SK Hynix Debuts on Nasdaq With $1 Trillion Valuation; Raises $29B for U.S. Expansion Breaking HalluSquatting attack tricks AI agents into running malicious code via hallucinated repos; 85% success rate Funding Europe VC funding hits $24B in Q2 2026—strongest quarter in 4 years as UK leads Breaking SpaceXAI releases Grok 4.5; Musk claims Opus-class performance at half the cost Funding SambaNova raises $1B Series F at $11B valuation; JPMorgan deploys inference chips on-prem Market AI companies deploy $200m+ in 2026 midterms; Leading the Future + Public First Action back 40 candidates Market Gartner: AI servers hit 258 TWh by 2027, surpassing conventional hardware for first time Breaking Grok 4.5 launches at $2/$6 per million tokens, undercutting Opus 4.8 on cost and speed Chips Imec roadmap: 2.5D optical I/O cuts AI inference interconnect power from 1.25kW to <200W Breaking SpaceXAI releases Grok 4.5 at $2/$6 per MTok; claims parity with Opus 4.7, faster

Research Thursday, July 9, 2026 at 01:04 PM

NVIDIA releases Nemotron Post-Training v3 Prompt Atlas—10T tokens, synthetic personas for agentic AI

NVIDIA released the Nemotron Post-Training v3 Prompt Atlas, an interactive visual map of over 10 trillion pre-training tokens and millions of post-training samples spanning multiple domains and agentic behaviors. The atlas lets researchers explore what's actually in Nemotron's training mixture: each point on the map represents a prompt sample, clustered by semantic similarity, and colorable by dataset, pipeline stage, domain, or tool use. Users can zoom into regions like coding algorithms, safety, math, and agentic behavior to inspect representative examples and understand model behavior.

The release includes Nemotron-Personas, locally grounded synthetic personas capturing regional demographic, geographic, and occupational diversity. Built using NVIDIA's NeMo Data Designer for compound-AI synthetic data generation, Nemotron-Personas now represents ten countries and more than 2.4 billion people. The personas are designed to help teams test whether their systems reflect the users, languages, regions, and occupations they claim to serve, with regional researchers and native speakers building quality into the data collaboratively.

NVIDIA's data strategy reflects a shift: open model weights matter, but for agents, reproducibility depends on datasets, curation choices, training recipes, and evaluation methods. Synthetic data provides a way to preserve useful signals without exposing proprietary sources, letting teams share richer signal without casually exposing underlying source data. Nemotron citations span 145 papers at ICML 2026, showing adoption across research.

For infrastructure and model teams, the atlas and persona library are tools for data inspection and evaluation. The emphasis on synthetic data and local quality control signals that future agent systems will depend on curated, domain-grounded training mixtures rather than undifferentiated scale. Teams building production agents should expect to invest in understanding their own training data mixture and regional grounding.

Source: huggingface.co →