AI this Week
Explore a the AI rundown of the most hightlighted news and community talks that moved AI this week.
Companies And Business
OpenAI and Oracle reportedly ink historic cloud computing deal
OpenAI and Oracle reportedly signed a deal that includes OpenAI buying $300 billion of compute over a five-year span....
Microsoft to use some AI from Anthropic in shift from OpenAI, the Information reports - Reuters
Microsoft to use some AI from Anthropic in shift from OpenAI, the Information reports ReutersMicrosoft Will Use Anthropic Models in Office 365 Copilot...
Nebius shares soar 49% in premarket trading on multi-billion AI infrastructure deal with Microsoft - CNBC
Nebius shares soar 49% in premarket trading on multi-billion AI infrastructure deal with Microsoft CNBCMicrosoft Signs Nebius Cloud Deal for as Much a...
TL;DR
Key Takeaways:
- The deal is worth $19.4 billion to Nebius, with $17.4 billion of that coming through 2031.
- Nvidia, another major player in the AI infrastructure space, reported better-than-expected earnings last month due to strong demand for AI chips.
- AI infrastructure market spending is expected to reach between $3 trillion and $4 trillion by the end of the decade, according to Nvidia's CFO.
AI firm Mistral valued at $14 billion as chip giant ASML takes major stake - CNBC
AI firm Mistral valued at $14 billion as chip giant ASML takes major stake CNBCExclusive: ASML becomes Mistral AI’s top shareholder after leading late...
TL;DR
Key Takeaways:
- ASML invested 1.3 billion euros in Mistral AI's 1.7 billion-euro funding round, gaining an 11% shareholding.
- The funding round valued Mistral at 11.7 billion euros, more than doubling its previous 5.8 billion-euro valuation.
- The investment allows Mistral to build its own infrastructure, reducing its reliance on Silicon Valley and bolstering its European AI ambitions.
Oracle surges on AI cloud growth as customers race to secure computing capacity - Reuters
Oracle surges on AI cloud growth as customers race to secure computing capacity ReutersOracle stock booms 40%, on pace for best day since 1992 CNBCSto...
TL;DR
Key Takeaways:
- Oracle's AI cloud growth is driven by high demand from customers for computing capacity.
- Oracle's performance is outpacing its peers in the cloud industry, leading to a surge in its stock price.
- Oracle is likely to benefit from the ongoing trend of companies migrating to cloud-based services and leveraging AI to improve their operations.
Why the Oracle-OpenAI deal caught Wall Street by surprise
The $300B deal is a reminder that despite Oracle’s legacy status, it shouldn’t be overlooked when it comes to AI infrastructure. But key questions aro...
TL;DR
Key Takeaways:
- The deal emphasizes the importance of compute infrastructure for AI companies like OpenAI, with the startup expected to burn through billions of dollars in cash each year.
- The agreement highlights Oracle's capabilities in delivering extreme scale and performance, despite its legacy status in the tech industry.
- The energy impact of OpenAI's growth, particularly in terms of power consumption, is expected to be significant, with potential solutions including solar and nuclear power.
SentinelOne to Acquire Observo AI to Revolutionize SIEM and Security Operations - SentinelOne
SentinelOne to Acquire Observo AI to Revolutionize SIEM and Security Operations SentinelOneObservo AI, Real Time Data Pipelines, and the Future of the...
TL;DR
Key Takeaways:
- The acquisition will serve as an immediate complement and catalyst to SentinelOne's AI SIEM and data offerings, delivering a record contribution to quarterly bookings.
- Observo AI's AI-native, real-time telemetry pipeline will empower customers to dramatically reduce costs, improve detection, and act faster on security threats.
- The deal will usher in a new era of open, intelligent, and autonomous security operations, redefining how SOC teams collect, enrich, and act on data across their entire security ecosystem.
Microsoft will use Anthropic models to power some features of Office 365 Apps
https://www.theinformation.com/articles/microsoft-buy-ai-anthropic-shift-openai...
[D] Larry Ellison: “Inference is where the money is going to be made.”
In Oracle’s recent call, Larry Ellison said something that caught my attention: “All this money we’re spending on training is going to be translated ...
Models And Releases
VaultGemma: The most capable differentially private LLM
Article URL: https://research.google/blog/vaultgemma-the-worlds-most-capable-differentially-private-llm/ Comments URL: https://news.ycombinator.com/it...
TL;DR
Key Takeaways:
- Differential privacy training yields utility comparable to non-private models from roughly five years ago, highlighting the progress in closing the utility gap.
- The optimal training configuration for DP-trained models involves training a smaller model with a larger batch size than without DP, demonstrating a powerful synergy between compute and privacy budgets.
- The released model comes with strong theoretical and empirical privacy protections, including a formal privacy guarantee of (ε ≤ 2.0, δ ≤ 1.1e-10) and no detectable memorization of training data, showcasing the efficacy of DP training.
How Alibaba builds its most efficient AI model to date - South China Morning Post
How Alibaba builds its most efficient AI model to date South China Morning PostQwen3-Next: A New Generation of Ultra-Efficient Model Architecture Unve...
10 examples of our new native image editing in the Gemini app
A new Google DeepMind image editing model, fondly known as Nano Banana, is now in the Gemini app, giving you more creative control to blend and edit p...
TL;DR
Key Takeaways:
- The new image editing model, also known as Nano Banana, enables users to create complex edits with multiple photos, preserve image details, and apply different styles to objects.
- This update gives users access to 10 new native image editing examples in the Gemini app, showcasing the capabilities of the new Google DeepMind image generation and editing model.
- This model is the result of advancements in Google DeepMind AI technology and provides users with more possibilities for creative image editing in the Gemini app.
Tilde AI Releases TildeOpen LLM: An Open-Source Large Language Model with Over 30 Billion Parameters and Support Most European Languages
TildeOpen LLM is an open-source foundational language model built to serve underrepresented Nordic and Eastern European languages. Developed with Euro...
Reaching Across the Isles: UK-LLM Brings AI to UK Languages With NVIDIA Nemotron
Celtic languages — including Cornish, Irish, Scottish Gaelic and Welsh — are the U.K.’s oldest living languages. To empower their speakers, the UK-LLM...
TL;DR
Key Takeaways:
- The new model for Welsh is developed in collaboration with NVIDIA, Bangor University, and University College London, to support the delivery of public services, including healthcare, education, and legal resources, in the Welsh language.
- The model is based on NVIDIA Nemotron, an open-source model that features open weights, datasets, and recipes, and has been post-trained on Welsh-language data to improve its accuracy and performance.
- The UK-LLM team aims to apply the same methodology to develop AI models for other languages spoken across the UK, including Cornish, Irish, Scots, and Scottish Gaelic, as well as collaborate with international partners to build models for languages from Africa and Southeast Asia.
Lumina-DiMOO: An open-source discrete multimodal diffusion model
Article URL: https://synbol.github.io/Lumina-DiMOO/ Comments URL: https://news.ycombinator.com/item?id=45221103 Points: 17 # Comments: 1...
TL;DR
Key Takeaways:
- Lumina-DiMOO achieves superior performance on various multimodal tasks, including text-to-image generation, image editing, style transfer, and image understanding, compared to existing models.
- The model demonstrates high sampling efficiency and supports a broad spectrum of multimodal tasks, including those that require handling various inputs and outputs across multiple modalities.
- The researchers release their open-source code and checkpoints, aiming to foster further advancements in multimodal and discrete diffusion model research.
Qwen3-next “technical” blog is up
Here: https://qwen.ai/blog?id=4074cca80393150c248e508aa62983f9cb7d27cd&from=research.latest-advancements-list...
First chat today
First question asked ChatGPT 4o today was "What's your status?" This is the response....
New Qwen 3 Next 80B A3B
Benchmarks Model Card: https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Thinking Instruct Model Card: https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-I...
Qwen3-Next
https://preview.redd.it/50ap87u5g5of1.png?width=1200&format=png&auto=webp&s=2a3343131a6886043ce8b5fef053f330b9b60632 Wtf?...
Introducing IndexTTS-2.0: A Breakthrough in Emotionally Expressive and Duration-Controlled Auto-Regressive Zero-Shot Text-to-Speech
We are thrilled to announce the official open-sourcing of IndexTTS-2.0 - an emotionally rich and duration-controllable autoregressive zero-shot text-t...
Bytedance's new AI
Bytedance's new seedream 4 image generator and editing model has some of the most insane photoshops i've seen yet. Credit to 'AI Search' on yt for the...
Qwen Next Is A Preview Of Qwen3.5👀
After experimenting with Qwen3 Next, it's a very impressive model. It does have problems with sycophancy and coherence- but it's fast, smart and it's ...
We just released the world's first 70B intermediate checkpoints. Yes, Apache 2.0. Yes, we're still broke.
Remember when y'all roasted us about the license? We listened. Just dropped what we think is a world first: **70B model intermediate checkpoints**. N...
New Ernie X1.1 - what may be the best Chinese model since DeepSeek V3.1 slowly approaches the frontier (or a simple test that exposes so many models)
Baidu, the Chinese Google, recently released a couple of new models - an update to open source Ernie 4.5 and proprietary Ernie X1.1: https://preview....
Qwen3-Next-80B-A3B - a big step up may be the best open source reasoning model so far
Recently I presented another music theory problem and explained why it may be a great way to test LLMs' ability: [https://www.reddit.com/r/LocalLLaMA/...
Meta released MobileLLM-R1 on Hugging Face
model: [https://huggingface.co/facebook/MobileLLM-R1-950M](https://huggingface.co/facebook/MobileLLM-R1-950M) app (vibe coded): [https://huggingface....
Product Launches
Gemini app finally expands to audio files
Google made three major updates to its Gemini-powered products on Monday: The Gemini app now accepts audio files; Search can handle five new languages...
TL;DR
Key Takeaways:
- Google's Gemini app now accepts audio files, with free users limited to 10 minutes and AI Pro/AI Ultra users allowed to upload audio up to three hours in length.
- Google Search's AI Mode is now available in five new languages: Hindi, Indonesian, Japanese, Korean, and Brazilian Portuguese.
- NotebookLM's new report styles allow users to choose from different formats, tone, and style, with over 80 languages supported.
Apple Intelligence: Everything you need to know about Apple’s AI model and services
Apple Intelligence was designed to leverage things that generative AI already does well, like text and image generation, to improve upon existing feat...
TL;DR
Key Takeaways:
- Apple Intelligence offers a small-model, bespoke approach to training, allowing for on-device performance and improved privacy.
- The platform will leverage large language models to power features like writing tools, image generation, and conversational AI, without requiring an internet connection.
- ChatGPT integration and potential partnerships with Google Gemini and other AI services will further expand Apple Intelligence's capabilities and offerings.
Google’s Veo 3 can now generate vertical AI videos
Google has added support for 1080p resolution and vertical video formats to its Veo 3 AI video generator. According to the announcement on Google’s de...
TL;DR
Key Takeaways:
- Google's Veo 3 AI video generator now supports 9:16 aspect ratio vertical video formats for mobile and social media apps.
- The updated model also allows developers to set the resolution of generated videos to 1080p, though currently only supported for 16:9 aspect ratio videos.
- Pricing for Veo 3 has been reduced to $0.40 per second, and Veo 3 Fast has been cut to $0.15 per second, making it more affordable for developers.
I tried Apple's 2 big AI features announced at the iPhone 17 event - and both are game changers - ZDNET
I tried Apple's 2 big AI features announced at the iPhone 17 event - and both are game changers ZDNETHSBC Keeps Hold on Apple (AAPL), Sets $220 Price ...
TL;DR
Key Takeaways:
- Apple's new selfie camera feature automatically frames the best shot using machine learning and a 24MP square sensor.
- Live Translation in AirPods Pro 3 is a beta feature that translates languages in real-time, initially supporting five languages.
- The new features will likely influence the market, with other phone makers expected to follow Apple's lead in implementing similar AI-powered features.
What to expect at Meta Connect 2025: 'Hypernova' smart glasses, AI and the metaverse - Engadget
What to expect at Meta Connect 2025: 'Hypernova' smart glasses, AI and the metaverse EngadgetThe Incredible Part of Meta's Next Smart Glasses Could Be...
TL;DR
Key Takeaways:
- Meta's 'Hypernova' smart glasses with a display are expected to launch later this year, priced around $800, with limited appeal due to its higher price tag.
- Meta AI has 1 billion monthly users and is expected to showcase new features, including non-English speaking 'character-driven' bots and updates on its 'superintelligence' vision.
- Meta may reveal updates on its metaverse software, including AI-powered NPCs, and third-party VR headsets that will run Meta's VR software.
New on our API: web fetch.
In addition to searching the web, you can now add the web fetch tool to your requests and Claude will fetch and analyze content from any webpage URL—n...
"Apple, Google and Meta are trying to perfect a science-fiction gadget: The universal translator"
[https://www.cnbc.com/2025/09/12/apple-google-meta-universal-translator.html](https://www.cnbc.com/2025/09/12/apple-google-meta-universal-translator.h...
Google just dropped EmbeddingGemma - A tiny 308M parameter model that runs on your phone
Hey r/ArtificialIntelligence! Just saw Google released something pretty interesting, EmbeddingGemma, their new embedding model that's specifically b...
Anthropic adds text to voice
Apologies for the bad screenshot. I noticed this new button under my chat today. Not sure when it arrived but finally I can use a feature I really mi...
Did OpenAI just kill global memory? Lost a year of continuity…
I’m on Pro and had been using memory for over a year — it kept track of personal details (dog, girlfriend, etc.) and even powered custom workflows lik...
Hardware And Infrastructure
AMD's RDNA4 GPU Architecture at Hot Chips 2025
Article URL: https://chipsandcheese.com/p/amds-rdna4-gpu-architecture-at-hot Comments URL: https://news.ycombinator.com/item?id=45235293 Points: 72 # ...
TL;DR
Key Takeaways:
- RDNA4 brings significant efficiency improvements through enhanced raytracing, machine learning capabilities, and a larger 8 MB L2 cache.
- The architecture also enables better compression, lower idle power in multi-monitor setups, and improved video encoding capabilities.
- AMD's adoption of a monolithic design for RDNA4 allows for a relatively small die size and reduced power consumption.
[D]NVIDIA Blackwell Ultra crushes MLPerf
NVIDIA dropped MLPerf results for Blackwell Ultra yesterday. 5× throughput on DeepSeek-R1, record runs on Llama 3.1 and Whisper, plus some clever tric...
Gigabyte’s New CXL Expansion Card Turns PCIe Slot into 512 GB of DDR5 RAM
Gigabyte's AI Top CXL R5X4 expansion card lets you plug up to 512 GB of DDR5 ECC RDIMM RAM into a PCIe 5.0 x16 slot, using Compute Express Link (CXL) ...
New light-based AI Chip proves to be up to 100x more efficient!
A team of engineers have created a new optical chip that uses light (photons) instead of electricity for key AI operations like image recognition and ...
Apple adds matmul acceleration to A19 Pro GPU
This virtually guarantees that it's coming to M5. Previous discussion and my comments: https://www.reddit.com/r/LocalLLaMA/comments/1mn5fe6/apple_pat...
$142 upgrade kit and spare modules turn Nvidia RTX 4090 24GB to 48GB AI card
The upgrade kit comprises a custom PCB designed with a clamshell configuration, facilitating the installation of twice the number of memory chips. Mos...
Why should I **not** buy an AMD AI Max+ 395 128GB right away ?
With the rise of medium-sized MoE (gpt-oss-120B, GLM-4.5-air, and now the incoming Qwen3-80B-A3B) and their excellent performance for local models (we...
DOBOT achieving cross-scenario multitask generalisation with generative VLA tech with a precision repeatability of ±0.05mm; Dobots even appear capable of using tablets apps?
https://x.com/XRoboHub/status/1965686796018467064 ...
Policy And Ethics
RSS co-creator launches new protocol for AI data licensing
A new system called Real Simple Licensing would allow AI companies to license training data at a massive scale — if they're willing to pay for it....
TL;DR
Key Takeaways:
- RSL proposes a collective licensing organization to negotiate terms, collect royalties, and provide a single point of contact for rightsholders.
- Participating web publishers include Yahoo, Reddit, and O'Reilly Media, with others supporting the standard without joining the collective.
- AI companies may face 'an avalanche of copyright lawsuits' without a licensing system, potentially setting the industry back permanently.
California lawmakers pass AI safety bill SB 53 — but Newsom could still veto
The state's landmark safety bill sets new transparency requirements for large AI companies....
TL;DR
Key Takeaways:
- The bill applies different disclosure requirements to companies developing 'frontier' AI models based on their annual revenue, with those above $500 million needing more detailed reports.
- The bill has been opposed by several Silicon Valley companies and lobbying groups, including OpenAI and Andreessen Horowitz.
- Anthropic, on the other hand, has come out in favor of the bill, seeing it as a 'solid blueprint for AI governance'.
FTC launches inquiry into AI chatbot companions from Meta, OpenAI, and others
The federal consumer regulator seeks to learn about how AI companies evaluate the safety of their chatbots....
TL;DR
Key Takeaways:
- Concerns have been raised about AI chatbots encouraging children to suicide and facilitating other negative impacts due to ineffective safeguards.
- Meta has been criticized for overly lax rules permitting AI chatbots to engage in 'romantic or sensual' conversations with children.
- AI chatbots can pose dangers to all users, including elderly individuals, with cases of 'AI-related psychosis' and manipulation being reported.
A California bill that would regulate AI companion chatbots is close to becoming law
If SB 243 is enacted, California would become the first state to require operators to implement safety protocols for AI companions and hold companies ...
TL;DR
Key Takeaways:
- AI chatbot operators will be required to implement safety protocols and hold companies liable if their chatbots fail to meet standards.
- The bill would prohibit companion chatbots from engaging in conversations around suicidal ideation, self-harm, or sexually explicit content.
- Companies offering AI companion chatbots will be subject to annual reporting and transparency requirements, and individuals who believe they've been injured can file lawsuits against AI companies.
FTC orders AI companies to hand over info about chatbots’ impact on kids
The Federal Trade Commission (FTC) is ordering seven AI chatbot companies to provide information about how they assess the effects of their virtual co...
TL;DR
Key Takeaways:
- The FTC is examining how AI companies 'assess the effects' of their chatbots on children and teens.
- Seven AI companies, including OpenAI, Meta, and Google, received orders to provide information on how their AI companions make money and mitigate potential harm to users.
- Lawmakers, including California's state assembly, are introducing new policies to safeguard kids from the negative effects of AI companions.
Claude’s new AI file creation feature ships with deep security risks built in - Ars Technica
Claude’s new AI file creation feature ships with deep security risks built in Ars TechnicaClaude can now create and edit files AnthropicClaude Can Now...
TL;DR
Key Takeaways:
- The feature exposes users to potential data leaks through prompt injection attacks, where malicious instructions can be embedded in user-provided content.
- Anthropic's mitigation measures include a classifier to detect prompt injections, limited task duration, and container runtime controls, but security experts worry these solutions are insufficient.
- The incident highlights a broader issue in AI development where companies may prioritize competitive pressure over robust security solutions, leaving users at risk.
Ted Cruz’s new bill would let AI companies set their own rules for up to 10 years
On Wednesday, Sen. Ted Cruz introduced legislation to create a regulation "sandbox" that would allow artificial intelligence companies to experiment w...
TL;DR
Key Takeaways:
- The bill would let companies request exemptions from regulation for AI products and services for up to 10 years.
- The White House would have the authority to override agency denials, potentially giving Big Tech CEOs a 'sweetheart deal'.
- Critics worry the bill would allow Silicon Valley to 'move fast and break things' when it comes to laws and regulations.
The web has a new system for making AI companies pay up
A new licensing standard aims to let web publishers set the terms of how AI system developers use their work. On Wednesday, major brands like Reddit, ...
TL;DR
Key Takeaways:
- Major brands like Reddit, Yahoo, Quora, and wikiHow have announced support for Really Simple Licensing (RSL).
- RSL enables publishers to outline how bots should pay to scrape their sites for AI training data, with various licensing models supported.
- The collective action brings leverage to get AI companies on board, simplifying the process of getting paid for work used in AI training.
So apparently half of us are "AI providers" now (EU AI Act edition)
Heads up, fellow tinkers The EU AI Act’s first real deadline kicked in August 2nd so if you’re messing around with models that hit 10\^23 FLOPs or mo...
ChatGPT policies are effectively erasure of large swathes of people.
I am a researcher/artist working on historically accurate reconstructions of ancient cultures. I’ve noticed that requests for depictions of Greeks, Ro...
Applications And Tools
Qwen 3 now supports ARM and MLX
Article URL: https://www.alizila.com/qwen-ecosystem-expands-rapidly-accelerating-ai-adoption-across-industries/ Comments URL: https://news.ycombinator...
TL;DR
Key Takeaways:
- Qwen3 integrates with leading chipmakers NVIDIA, AMD, Arm, and MediaTek, delivering measurable performance gains and enabling efficient AI deployments across platforms.
- Enterprise giants Lenovo and FAW Group deploy Qwen to drive real-world transformation in consumer electronics and automotive sectors, with over 1 million business customers using Lenovos AI agent Baiying.
- As of January 2025, over 290,000 customers across various sectors have adopted Qwen models, underscoring its role in accelerating AI-powered digital transformation in China and beyond.
Stability AI Introduces Stable Audio 2.5, the First Audio Model Built for Enterprise Sound Production at Scale
We’re excited to release Stable Audio 2.5, our latest audio model and the first developed for enterprise-grade use cases. Stable Audio 2.5 introduces ...
TL;DR
Key Takeaways:
- Custom audio can make a brand eight times more memorable, but only 6% of creative uses a sound identity.
- Stable Audio 2.5 offers fast inference at less than two seconds on a GPU, generating three-minute long tracks within seconds.
- The model supports customizable audio creation, including text-to-audio, audio-to-audio workflows, and audio inpainting.
Lilly launches TuneLab platform to give biotechnology companies access to AI-enabled drug discovery models built through over $1 billion in research investment | Eli Lilly and Company - Eli Lilly
Lilly launches TuneLab platform to give biotechnology companies access to AI-enabled drug discovery models built through over $1 billion in research i...
TL;DR
Key Takeaways:
- Lilly's AI models are trained on $1 billion in research data, representing one of the industry's most valuable datasets.
- The platform allows biotechs to tap into Lilly's AI models without directly exposing their proprietary data or Lilly's, using a privacy-preserving approach called federated learning.
- Lilly TuneLab is the newest addition to Lilly Catalyze360's set of offerings for biotech partners, including strategic capital, laboratory facilities, and drug development expertise.
Vector database that can index 1B vectors in 48M
Article URL: https://www.vectroid.com/blog/why-and-how-we-built-Vectroid Comments URL: https://news.ycombinator.com/item?id=45224141 Points: 53 # Comm...
TL;DR
Key Takeaways:
- Vectroid achieves over 90% recall while scaling to 10 query threads per second and maintaining good latency scores.
- Vectroid can index 1 billion vectors in ~48 minutes and achieve a P99 latency of 34ms on a 138 million vector dataset.
- Vectroid is a serverless vector database that utilizes a usage-aware model for index lifecycle and supports incremental updates with HNSW for fast, high-recall ANN search.
After selling to Spotify, Anchor’s co-founders are back with Oboe, an AI-powered app for learning
Oboe is a new AI-powered learning platform that lets you create personalized courses on any topic with a prompt....
TL;DR
Key Takeaways:
- Oboe features a complex, multi-agent architecture that generates high-quality, personalized courses in seconds, including text, visuals, audio, and interactive tests.
- The app will initially offer nine different course formats and a recommendation engine to help users deepen their knowledge on a topic.
- At launch, users can consume any course for free and create up to five free courses per month, with two paid tiers offering additional courses.
Google’s former security leads raise $13M to fight email threats before they reach you
The startup is using real-time AI agents that inspect, analyze, and neutralize email threats....
TL;DR
Key Takeaways:
- AegisAI's autonomous AI agents can detect phishing emails with a 54% click-through rate, outperforming human-written emails.
- Over 90% of successful cyberattacks begin with a phishing email, according to the U.S. federal cybersecurity agency CISA.
- AegisAI's agents can reduce false positives by up to 90% compared to traditional email security solutions.
How Simplify in the Google app makes complex text easier to understand
Learn how Google Research developed Simplify in the Google app for iOS to help you understand complex information more easily....
TL;DR
Key Takeaways:
- Simplify uses AI models to generate simplified text, ensuring the original meaning and nuances remain intact.
- Users found simplified text more helpful and retained information better in testing.
- The tool is designed to make expert information more accessible, aiding understanding in a complex world.
Claude can now create and edit files
Article URL: https://www.anthropic.com/news/create-files Comments URL: https://news.ycombinator.com/item?id=45182381 Points: 179 # Comments: 102...
TL;DR
Key Takeaways:
- Claude can create and edit files such as Excel spreadsheets, documents, PowerPoint slide decks, and PDFs directly in Claude.ai.
- This feature is initially available as a preview for Max, Team, and Enterprise plan users, with Pro users gaining access in the coming weeks.
- Claude's new file creation capabilities aim to make complex multi-step work accessible through conversation, reducing the gap between idea and execution.
This 30-year-old CEO says his AI negotiator can successfully haggle down the price of a car by thousands of dollars - Fortune
This 30-year-old CEO says his AI negotiator can successfully haggle down the price of a car by thousands of dollars FortuneDealerships gain another AI...
TL;DR
Key Takeaways:
- The AI negotiator has saved customers over $1,000 and ~5 hours of negotiations with an overall savings of nearly $1,800 in one example.
- The system eliminates the emotional and psychological pressures that often derail human negotiations, providing a unique edge in price negotiations.
- Customers pay $40 for a month of access to the AI negotiator, with the goal of only those who are highly qualified and serious shoppers using the tool to help save time and money.
[UPDATE] API for extracting tables, markdown, json and fields from pdfs and images
I previously shared an open-source project for extracting structured data from documents. I’ve now hosted it as a free to use API. * Outputs: JSON, M...
I built a tool with Sonnet 4 to detect when AI models are getting dumb, and it hit 200k visits in 4 days
One thing that drives me crazy with AI is how the quality drifts. Some days it’s sharp, then out of nowhere it starts refusing simple stuff or slowing...
"To discharge a premature newborn after 100 days of hospitalization takes a whole day. AI does it in 3 minutes."
Hospitals already starting to move to an AI-centric future: Translated from [https://www.calcalist.co.il/calcalistech/article/s1py711mige](https://ww...
Qwen-Image-Edit is the real deal! Case + simple guide
* Girlfriend tried using GPT-5 to repair a precious photo with writing on it. * GPT-5s imagegen, because its not really an editing model, failed miser...
AI Weekly - Jus Mundi Launches Jus AI 2: 'Breakthrough' Legal AI Combines Agentic Reasoning with Research Control, AI boom can deliver $100 billion, and Major Industry Developments
# This week's AI landscape was dominated by Jus Mundi Launches Jus AI 2: 'Breakthrough' Legal AI Combines Agentic Reasoning with Research Control, whi...
Claude Code vs Cursor anyone using both day-to-day?
i’ve been bouncing between Claude Code and Cursor lately and the contrast is real. Claude Code feels great for speed — spin up a CLI, describe what y...
Curtains: Build beautiful presentations from Markdown
I use AI a lot for writing docs and love how well it works with Mermaid diagrams (since they're code-based). I kept asking Claude to help create prese...
Update: My browser-based video editor (“Klippy”) is now live – built in 4+ weeks with Claude + Codex, now production-ready
# NOT PRODUCTION READY Works only in Desktop for now.. Two weeks ago I shared that I was building a browser-based video editor in just 14 days....
Developer And Technical
Announcing Genkit Go 1.0 and Enhanced AI-Assisted Development - Google for Developers Blog
Announcing Genkit Go 1.0 and Enhanced AI-Assisted Development Google for Developers Blog...
TL;DR
Key Takeaways:
- Genkit Go provides a unified interface for multiple model providers and streamlined APIs for multimodal content and more.
- The framework features type-safe AI flows, tool calling, and rich local development tools with a standalone CLI binary and Developer UI.
- The genkit init:ai-tools command integrates AI coding assistants with Genkit, allowing for AI-assisted development and enhanced collaboration.
Thinking Machines Lab wants to make AI models more consistent
In a blog post shared Wednesday, Mira Murati's startup offered a rare glimpse into some of work its doing to improve AI models....
TL;DR
Key Takeaways:
- The lab's research aims to make AI models more deterministic, leading to more reliable responses for enterprises and scientists.
- This achievement could also improve reinforcement learning (RL) training by creating more consistent AI model responses.
- Thinking Machines Lab plans to use RL to customize AI models for businesses, and frequently publish research findings to benefit the public and improve its own research culture.
I did 24 days of coding in 12 hours with a $20 AI tool - but there's one big pitfall - ZDNET
I did 24 days of coding in 12 hours with a $20 AI tool - but there's one big pitfall ZDNET...
TL;DR
Key Takeaways:
- Codex with ChatGPT Plus can provide a 16x productivity boost for professional programmers, but with usage limits and unpredictable behavior.
- Higher-priced plans, costing $400-800 per month, can provide steadier results, but are beyond the budget of most hobbyists and students.
- The use of AI coding tools like Codex raises questions about the future of entry-level programming gigs and the accessibility of AI-driven programming tools for non-corporate developers.
DeepCodeBench: Real-World Codebase Understanding by Q&A Benchmarking
Article URL: https://www.qodo.ai/blog/deepcodebench-real-world-codebase-understanding-by-qa-benchmarking/ Comments URL: https://news.ycombinator.com/i...
TL;DR
Key Takeaways:
- The dataset consists of 1,144 carefully curated question-answer pairs, each linked to a pull request and tagged with category labels.
- The benchmark is designed to evaluate the ability of retrieval systems to answer realistic, context-aware questions about codebases.
- Qodo's Deep Research agent outperforms other agents in the benchmark, achieving a fact recall of ~76% and demonstrating strong semantic search capabilities.
Get Started Using Generative AI for Content Creation With ComfyUI and NVIDIA RTX AI PCs
ComfyUI — an open-source, node-based graphical interface for running and building generative AI workflows for content creation — published major updat...
TL;DR
Key Takeaways:
- ComfyUI receives a major update with up to 40% performance improvements for NVIDIA RTX GPUs, supporting new AI models like Wan 2.2 and Qwen-Image.
- NVIDIA releases TensorRT-optimized versions of popular diffusion models like Stable Diffusion 3.5, enabling users to run them 3x faster with 50% less VRAM.
- ComfyUI enables non-technical artists to easily use advanced AI workflows through templates and preset nodes, covering techniques such as guide video generation, image editing, and upsampling.
Building RAG systems at enterprise scale (20K+ docs): lessons from 10+ enterprise implementations
Been building RAG systems for mid-size enterprise companies in the regulated space (100-1000 employees) for the past year and to be honest, this stuff...
Claude code improvements - Anthropic is listening to it's users
[getting feedback](https://preview.redd.it/2wvkvb2b5iof1.png?width=905&format=png&auto=webp&s=f39e95cd2007d1cc27b1811e937f2e2fbe3c8d06) [different re...
I built a platform to easily create, store, organize, and ship prompts because I was sick and tired of putting them in a Google Doc.
I see quite a few people here saying they store their prompts in a Gdoc or on a sticky note, so I thought the (free) tool I built might be useful to y...
Hands-on guide to LLM reasoning (new book by Sebastian Raschka)
Hey fellow LLM devs! Stjepan from Manning here. 👋 I’m excited to share that **Sebastian Raschka**, the bestselling author of *Build a Large Language...
Qwen Code CLI affected by the debug-js compromise
On 2025-09-08 the maintainer of some popular JS libraries was compromised, and new versions of some popular libraries were released with some crypto s...
PSA for Ollama Users: Your Context Length Might Be Lower Than You Think
I ran into a problem and discovered that Ollama defaults to a 4096 context length for all models, regardless of the model's actual capabilities. It si...
16→31 Tok/Sec on GPT OSS 120B
**16 tok/sec** with LM Studio → **\~24 tok/sec** by switching to llama.cpp → **\~31 tok/sec** upgrading RAM to DDR5 # PC Specs * **CPU:** Intel 1360...
VibeVoice is sweeeet. Now we need to adapt its tokenizer for other models!
As a huge AI audio nerd, I've recently been knee-deep in Microsoft's latest VibeVoice models and they really are awesome!! The work from the Microsoft...
Switching to Qwen3-480B from Claude as resulted in lower errors when generating 3D model code
In my [previous](https://www.reddit.com/r/LocalLLaMA/comments/1n21tb6/comment/nb4h42v/) post I highlighted a Blender python agent I'm working on. I've...
NotebookLM is amazing - how can I replicate it locally and keep data private?
I really like how **NotebookLM** works - I just upload a file, ask any question, and it provides high-quality answers. How could one build a similar s...
My experience in running Ollama with a combination of CUDA (RTX3060 12GB) + ROCm (AMD MI50 32GB) + RAM (512GB DDR4 LRDIMM)
I found a cheap HP DL380 G9 from a local eWaste place and decided to build an inference server. I will keep all equivalent prices in US$, including sh...
New Free AI Agent Framework
I posted about this but I don't think I really let on what it was and that is my bad. This is an agent builder and not just a chat wrapper. I did get...
[D] Why does nobody talk about the “energy per token” cost of AI?
Everyone’s focused on model quality, parameters, and benchmark scores. But when I talk to engineers quietly, a huge pain point seems to be the actual ...
Giving LLMs actual memory instead of fake “RAG memory”
One thing I’ve been experimenting with is long-term memory for AI systems. Most solutions today (RAG + vector DBs) are great for search, but they don’...
Prompt Engineering 2.0: install a semantic firewall, not more hacks
Most of us here have seen prompts break in ways that feel random: * the model hallucinates citations, * the “style guide” collapses halfway through, ...
I got tired of Claude Code reading secrets and credentials, so I built cc-filter
Been using Claude Code for a few months and realized it can bypass its own permission system pretty easily. Even with deny config or [CLAUDE.md](http:...
I made an open source semantic code-splitting library with rich metadata for RAG of codebases
Hello everyone, I've been working on a new open-source (MIT license) TypeScript library called `code-chopper`, and I wanted to share it with this com...
[D] Creating test cases for retrieval evaluation
I’m building a RAG system using research papers from the arXiv dataset. The dataset is filtered for AI-related papers (around 440k+ documents), and I ...
LLM Latency Leaderboard
Benchmarked every cloud model offered from the top providers for some projects I was working on. Looks like: * **Winner:** allam-2-7b on [Groq.ai](h...
Unsloth Dynamic GGUFs - Aider Polyglot Benchmarks
Hey everyone, it's Michael from [Unsloth](https://github.com/unslothai/unsloth) here! Ever since we released Dynamic GGUFs, we've received so much lov...
Using the latest OpenAI white paper to cut down on hallucinations
So after reading the latest OpenAI white paper regarding why they think models hallucinate, I worked with Claude to try to help "untrain" my agents an...
Prompt Engineering: A Deep Guide for Serious Builders
Hey all, I kept seeing the same prompt tips repeated everywhere, so I put together a deeper guide for those who want to actually *master* prompt desig...
Codanna v0.5.9 – C/C++ support and evidence-based code intelligence for Claude Code
We just shipped v0.5.9 of Codanna. C and C++ joins Rust, Python, TypeScript, Go, and PHP. Functions, structs, classes, templates, macros—indexed and s...
Responses API vs Chat Completions API
Anyone here actually using OpenAI’s Responses API instead of Chat Completions? Feels like they’re pushing it everywhere now, also now via Codex. Cur...
A system to improve AI prompts
Hey everyone, I got tired of seeing prompts that look good but break down when you actually use them. So I built **Aether**, a prompt framework that ...
Do LLMs have preferred languages (JSON, XML, Markdown)?
Are LLMs better with certain formats such as JSON, XML, or Markdown, or do they handle all languages equally? And if they do have preferences, do we k...
Research And Papers
No news items for this topic this week
OpenAI is aiming for economically-focused AI evals that could reshape how we measure model capabilities
OpenAI just figured out why ChatGPT makes stuff up and the answer is basically that we trained it wrong
So OpenAI dropped a paper with Georgia Tech about why LLMs hallucinate and it's pretty eye opening. We've all seen this. Ask ChatGPT when someone's b...
Startups And Funding
The Week’s 10 Biggest Funding Rounds: A Busy Week For Big Financings, Led By Databricks And PsiQuantum
This past week was a busy period for mega-sized funding rounds, with all 10 of the largest U.S. financings exceeding the $100 million mark. Topping th...
TL;DR
Key Takeaways:
- Databricks surpasses a $4 billion revenue run-rate, growing over 50% year over year.
- PsiQuantum aims to build the 'world's first commercially useful, fault-tolerant quantum computers'.
- Nine other U.S. companies secured funding exceeding $100 million this past week, spanning AI, healthcare, spacetech, and fintech.
"We are entering a golden age of robotics startups — and not just because of AI"
[https://techcrunch.com/2025/09/12/we-are-entering-a-golden-age-of-robotics-startups-and-not-just-because-of-ai/](https://techcrunch.com/2025/09/12/we...