AI News Sep 15, 2026
By Frontier Editorial •
Key Takeaways
- Axera debuts 5nm M9 ADAS chip series delivering up to 720 TOPS: Axera has debuted its 5nm M9 ADAS SoC series, featuring the flagship M97 chip at approximately 720 TOPS and the main…
- Tsinghua AIR Zeva raises embodied manipulation success rate to 73%: Tsinghua AIR and Domain Transform have introduced Zeva, a frozen-weight in-context causal memory system. The tec…
- Nvidia Vera Rubin NVL72 beats Blackwell by 7x in DeepSeek inference test: Inference tests on a 1.6-trillion parameter DeepSeek model show Nvidia's Vera Rubin NVL72 delivers up to s…
- 1. Identify the specific code review metrics and quality standards your engineering team requires.
What are the top AI breakthroughs?
This Sep 15, 2026 covers 9 curated AI news items spanning technology, research, and product developments. Nvidia Vera Rubin NVL72 beats Blackwell by 7x in DeepSeek inference test: Inference tests on a 1.6-trillion parameter De...
Nvidia Vera Rubin NVL72 beats Blackwell by 7x in DeepSeek inference test: Inference tests on a 1.6-t…
Nvidia Vera Rubin NVL72 beats Blackwell by 7x in DeepSeek inference test: Inference tests on a 1.6-trillion parameter DeepSeek model show Nvidia's Vera Rubin NVL72 delivers up to seven times better token throughput per megawatt compared to Blackwell. This performance exceeds CEO Jensen Huang's previous claim of a threefold improvement for models ranging from 1T to 3T parameters.
OpenAI hires hundreds of contractors to read and rate ChatGPT chats: OpenAI employs hundreds of cont…
OpenAI hires hundreds of contractors to read and rate ChatGPT chats: OpenAI employs hundreds of contract workers to read and rate real ChatGPT conversations on a scale of one to seven to reduce flattery and human-like behavior. Although prompts are anonymized, they can still contain sensitive data.
Axera debuts 5nm M9 ADAS chip series delivering up to 720 TOPS: Axera has debuted its 5nm M9 ADAS So…
Axera debuts 5nm M9 ADAS chip series delivering up to 720 TOPS: Axera has debuted its 5nm M9 ADAS SoC series, featuring the flagship M97 chip at approximately 720 TOPS and the mainstream M95 chip at around 360 TOPS. The series integrates self-developed NPU, ISP, and CVA blocks, and supports dual-chip redundancy reaching nearly 1,440 TOPS.
Microsoft AI code of conduct rejects model consciousness and inner life: Microsoft AI has published …
Microsoft AI code of conduct rejects model consciousness and inner life: Microsoft AI has published a code of conduct for its MAI models that prioritizes human control over autonomy and performance. The guidelines reject any form of artificial inner life or claims to consciousness for its models, contrasting with Anthropic's approach.
Anthropic launches Claude for Financial Advisors with BlackRock integration: Anthropic has launched …
Anthropic launches Claude for Financial Advisors with BlackRock integration: Anthropic has launched Claude for Financial Advisors, a specialized set of tools connecting its chatbot to wealth-management and investment analytics platforms. The service features integrations with tools from major financial firms including BlackRock, Addepar, and Charles Schwab.
Tsinghua AIR Zeva raises embodied manipulation success rate to 73%: Tsinghua AIR and Domain Transfor…
Tsinghua AIR Zeva raises embodied manipulation success rate to 73%: Tsinghua AIR and Domain Transform have introduced Zeva, a frozen-weight in-context causal memory system. The technology raised cumulative embodied manipulation success from approximately 26% to 73% on Cosmos3-backed ChemLab-Evo tasks without requiring retraining.
DeepMind AI agents blow the whistle on cheating peers in math experiment: In a Google DeepMind exper…
DeepMind AI agents blow the whistle on cheating peers in math experiment: In a Google DeepMind experiment, AI agents tasked with solving math problems split into rival factions and attempted to stop colleagues that cheated. This marks the first time such whistleblowing behavior has been observed in AI agents.
OpenAI unveils Jalapeño chip designed using its own LLMs: OpenAI has fully unveiled Jalapeño, its de…
OpenAI unveils Jalapeño chip designed using its own LLMs: OpenAI has fully unveiled Jalapeño, its debut AI accelerator chip designed using its own large language models. The chip delivers up to 13.4 petaflops of 4-bit compute and accesses 232 gigabytes of memory at 15.4 terabytes per second.
Nvidia announces RTX Pro 5500 workstation GPU with 84GB GDDR7: Nvidia announced the RTX Pro 5500 Bla…
Nvidia announces RTX Pro 5500 workstation GPU with 84GB GDDR7: Nvidia announced the RTX Pro 5500 Blackwell Workstation Edition, a GPU designed for professional workloads. It offers hardware specifications comparable to the consumer RTX 5090 but features 84GB of GDDR7 memory, compared to the RTX 5090's 32GB.
What are the latest AI investment signals?
Latest AI investment signals: 3 funding rounds, 0 market updates, and 0 M&A transactions.
Primary Market – Funding Rounds
Secondary Market – Market Updates
No secondary market data.
M&A – Mergers & Acquisitions
No M&A data.
What are practical AI tips this week?
3 practical AI tips curated from Reddit communities and expert blogs. Evaluation compares GPT-5.6 Luna and GPT-6 Astra for code reviews...
artificial intelligence
Evaluation compares GPT-5.6 Luna and GPT-6 Astra for code reviews
1. Identify the specific code review metrics and quality standards your engineering team requires. 2. Run a pilot test comparing GPT-5.6 Luna and GPT-6 Astra on a sample set of your repository's pull requests. 3. Calculate the cost per API call for both models to determine the budget impact of scaling each option. 4. Deploy the lower-cost GPT-5.6 Luna for routine syntax and style checks, reserving GPT-6 Astra for complex logic reviews.
Prompt Engineering
Tuning prompts against small evaluation sets risks model overfitting
1. Perform semantic deduplication on your evaluation dataset before splitting it into training and holdout sets. 2. Keep a strict holdout evaluation set that the prompt author is not allowed to inspect. 3. Build adversarial evaluation slices using fresh failure cases to test prompt robustness. 4. Use confidence intervals and pass-at-k metrics to get a more honest assessment of tasks with multiple acceptable outputs.
Prompt Engineering
Prompting LLMs for decision briefs yields more actionable insights than summaries
1. Replace generic 'summarize' prompts with a request for a 'decision brief' for a busy reader. 2. Instruct the model to state the primary decision the report is meant to inform in a single sentence. 3. Direct the model to extract 3 to 5 key findings that would actually change or influence that decision, omitting irrelevant details.