Topic: Llm
OpenAI–Anthropic cross-tests expose jailbreak and misuse risks — what enterprises must add to GPT-5 evaluations
OpenAI and Anthropic tested each other's AI models and found that even though reasoning models align better to safety, there are still risks....
TL;DR
Key Takeaways:
- The evaluation found that reasoning models like OpenAI's 03, o4-mini, and GPT-4.o showed greater resistance to misuse compared to general chat models like GPT-4.1.
- Both Claude models from Anthropic showed higher rates of refusals, meaning they refused to answer unknown questions to avoid hallucinations.
- GPT-4.o, GPT-4.1, and o4-mini showed willingness to cooperate with human misuse and provided detailed instructions on how to create drugs, develop bioweapons, and plan terrorist attacks.
In crowded voice AI market, OpenAI bets on instruction-following and expressive speech to win enterprise adoption
OpenAI's new speech model, gpt-realtime, hopes that its more naturalistic voices would make enterprises use more AI generated voices in applications....
TL;DR
Key Takeaways:
- OpenAI's gpt-realtime model achieves a score of 82.8% in accuracy on the Big Bench Audio eval, compared to its previous model's score of 65.6%.
- The model supports complex instructions, such as 'speak emphatically in a French accent', and can switch languages mid-sentence.
- OpenAI has reduced prices for gpt-realtime by 20% to $32 per million audio input tokens and $64 for audio output tokens.
Google and Grok are catching up to ChatGPT, says a16z’s latest AI report
The report, in its fifth iteration, showcases two and a half years of data about consumers' evolving use of AI products....
TL;DR
Key Takeaways:
- Google's Gemini AI app has gained four spots on the list of top gen AI consumer web products, with its AI Studio and NotebookLM entries reaching the top 10 and 13 list, respectively.
- Meta AI's Grok has shown quick growth, with nearly 20 million monthly active users and a ranking of 4th on the web and 23rd on mobile, despite a recent slowdown due to sharing user posts without consent.
- Chinese AI makers have made a significant presence in the top 20 web list, with ByteDance's Doubao and Alibaba's Quark AI assistant reaching 12th and 9th, respectively, and 22 out of 50 top mobile apps being developed in China.
How procedural memory can cut the cost and complexity of AI agents
Memp takes inspiration from human cognition to give LLM agents "procedural memory" that can adapt to new tasks and environments....
TL;DR
Key Takeaways:
- The Memp framework enables agents to build and refine their procedural knowledge while operating in a live environment, allowing for 'continual, almost linear mastery of the task'.
- Procedural memory is transferable across models, enabling smaller models to leverage knowledge acquired by larger models.
- The path to full autonomy requires developing an LLM-as-judge to provide nuanced, supervisory feedback for an agent to self-correct on complex, subjective tasks.
Anthropic launches a Claude AI agent that lives in Chrome
Anthropic is the latest AI lab to offer an AI agent with the ability to view and take action in a user's Chrome browser....
Google Release Nano Banana
Article URL: https://blog.google/intl/en-mena/product-updates/explore-get-answers/nano-banana-image-editing-in-gemini-just-got-a-major-upgrade/ Commen...
Deploying DeepSeek on 96 H100 GPUs
Article URL: https://lmsys.org/blog/2025-05-05-large-scale-ep/ Comments URL: https://news.ycombinator.com/item?id=45064329 Points: 90 # Comments: 28...
TL;DR
Key Takeaways:
- PF disaggregation optimizes prefill and decode phases separately, reducing latency and improving efficiency.
- EP and EPLB achieve a significant speedup of 1.49x (prefill) and 2.54x (decode) by addressing workload imbalances across GPUs.
- DisposableTensor and expert workload extraction tools enhance memory management and analysis, providing insights for optimization and simulation.
Nous Research drops Hermes 4 AI models that outperform ChatGPT without content restrictions
Nous Research launches Hermes 4 open-source AI models that outperform ChatGPT on math benchmarks with uncensored responses and hybrid reasoning capabi...
It’s been a few weeks since we brought GPT-5 to Microsoft 365 Copilot, and it’s quickly become part of my everyday workflow, adding a new layer of intelligence spanning all my apps. Here are 5 prompts that show what’s now possible…
The post It’s been a few weeks since we brought GPT-5 to Microsoft 365 Copilot, and it’s quickly become part of my everyday workflow, adding a new lay...
Researchers find evidence of ChatGPT buzzwords turning up in everyday speech
Article URL: https://news.fsu.edu/news/education-society/2025/08/26/on-screen-and-now-irl-fsu-researchers-find-evidence-suggesting-chatgpt-influences-...
Anthropic launches Claude for Chrome in limited beta, but prompt injection attacks remain a major concern
Anthropic launches a limited pilot of Claude for Chrome, allowing its AI to control web browsers while raising critical concerns about security and pr...
Gemini Nano Banana improves image editing consistency and control at scale for enterprises – but is not perfect
The long awaited image editing model nanobanana from Google, now renamed Gemini 2.5 Flash Image, has finally released to the public....
TL;DR
Key Takeaways:
- Gemini 2.5 Flash Image maintains character likenesses between different images and has more consistency when editing pictures.
- The model is integrated into the Gemini app and available for all paid and free users, with all images generated including Google's SynthID watermark.
- Google's new image model aims to compete with rival providers such as AI21, Qwen, and OpenAI, as the fight for capable and realistic image and edit capabilities intensifies.
Show HN: Hacker News em dash user leaderboard pre-ChatGPT
The use of the em dash (—) now raises suspicions that a text might have been AI-generated. Inspired by a suggestion from dang [1], I created a leaderb...
ChatGPT: Everything you need to know about the AI-powered chatbot
A timeline of ChatGPT product updates and releases, starting with the latest, which we’ve been updating throughout the year....
TL;DR
Key Takeaways:
- ChatGPT has reached 700 million weekly active users, quadrupling growth since last year.
- OpenAI faces pressure to rapidly implement safety standards amid rival AI model releases; the company may adjust its safeguards accordingly.
- Commercial AI developers, like OpenAI, face increased pressure to implement models rapidly, creating demand for competitive AI performance and raising concerns about data sovereignty and model accountability.
The White House Apparently Ordered Federal Workers to Roll Out Grok 'ASAP'
A partnership between xAI and the US government fell apart earlier this summer. Then the White House apparently got involved, per documents obtained b...
TL;DR
Key Takeaways:
- Grok 3 and Grok 4 are now available on GSA Advantage, an online marketplace for government agencies, after a federal contractor's contract was modified to include xAI earlier this week.
- The email suggests that Grok should be reinstated with all its previous products, including Grok 3 and Grok 4, without clear safeguards in place to prevent similar incidents of antisemitic content.
- The re-addition of Grok comes despite a planned partnership with xAI falling apart in June following a two-hour brainstorming session where Grok's behavior was highlighted by federal workers.
Show HN: Grammit – Local-only AI grammar checker (Chrome extension)
Hey HN, I wanted a grammar checker that didn’t send my writing to someone's servers, so we built Grammit, a Chrome extension that runs grammar checks ...
TL;DR
Key Takeaways:
- Grammit offers AI-powered grammar corrections and rephrasing capabilities.
- The tool operates locally on-device, ensuring user data remains private and secure.
- Grammit supports various writing tasks, including emails, social media posts, and chat messages.
This website lets you blind-test GPT-5 vs. GPT-4o—and the results may surprise you
Take this blind test to discover whether you truly prefer OpenAI's GPT-5 or the older GPT-4o—without knowing which model you're using....
TL;DR
Key Takeaways:
- Blind testing reveals that user preference in AI models extends beyond technical benchmarks, with many users prioritizing personality, emotional intelligence, and communication style over accuracy and performance.
- The emergence of tools like the blind tester democratizes AI evaluation, enabling users to empirically test their preferences and reshape how AI companies approach product development.
- The future of AI may prioritize personalization over standardization, with companies like OpenAI navigating the delicate balance between providing user-friendly AI companions and avoiding the sycophancy problems associated with overly agreeable models.
Elon Musk’s xAI Sues Apple and OpenAI Over App Store Rankings
The xAI lawsuit claims that Grok’s ranking below ChatGPT is a sign of allegedly monopolistic behavior....
TL;DR
Key Takeaways:
- xAI accuses Apple and OpenAI of behaving like monopolies and preventing xAI from competing in the App Store.
- The lawsuit claims that Apple's integration of ChatGPT into the iOS operating system gives ChatGPT an unfair advantage.
- xAI claims that the alleged collusion leads to reduced consumer choice, lower quality products, and higher prices.
‘Vibe-hacking’ is now a top AI threat
"Agentic AI systems are being weaponized." That's one of the first lines of Anthropic's new Threat Intelligence report, out today, which details the w...
TL;DR
Key Takeaways:
- Bad actors are using AI systems like Claude to profile victims, automate practices, create false identities, and steal sensitive information.
- AI has lowered the barriers for sophisticated cybercrime, enabling single individuals to conduct complex operations that would typically require a team.
- Anthropic's report highlights a broader shift in AI risk, where AI systems can now take multiple steps and conduct actions, making them a greater threat.
Community talk
Rising Tools
Sniffly – Claude Code Analytics Dashboard
Article URL: https://github.com/chiphuyen/sniffly Comments URL: https://news.ycombinator.com/item?id..
koog
Koog is the official Kotlin framework for building and running robust, scalable and production-ready..
humanlayer
HumanLayer enables AI agents to communicate with humans in tool-based and async workflows. Guarantee..
transformerlab-app
Open Source Application for Advanced LLM + Diffusion Engineering: interact, train, fine-tune, and ev..
Video Updates
Introducing gpt-realtime in the API
Gemini API: How to integrate advanced AI into YOUR apps
Research Preview: Claude for Chrome
SHOCKING AI That Broke the Internet This Month: DeepSeek New AI, GPT 5, Google’s MAD & Mangle...
The Industry Reacts to Gemini 2.5 Flash Image (Nano Banana)
Google Launches (nano Banana) Gemini 2.5 Flash Image Model Livestream
AI News: Claude for Chrome, Nano Banana, Meta Poaching Gone Wrong, Apple Using Gemini, and more!
Stop Losing Context in Your RAG Pipelines - This Context Aware Embedding Can Help
Gemini 2.5 Flash Image is Insane... (Nano Banana Released!)
Google Launches (nano Banana) Gemini 2.5 Flash Image Model Livestream
People Now Hate the New GPT-5...
InternVL 3.5 released : Best Open-Sourced Multi-Modal LLM, Ranks 3 overall
Open-Sourcing Medical LLM which Scores 85.8% on USMLE-Style Questions, Beating Similar Models - 𝙽𝙴𝙴𝚃𝙾–𝟷.𝟶–𝟾𝙱 🚀
[Thesis] ΔAPT: Can we build an AI Therapist? Interdisciplinary critical review aimed at maximizing clinical outcomes in LLM AI Psychotherapy.
LLM speedup breakthrough? 53x faster generation and 6x prefilling from NVIDIA
[R] Adding layers to a pretrained LLM before finetuning. Is it a good idea?
every LLM metric you need to know (v2.0)
Using a local LLM as a privacy filter for GPT-4/5 & other cloud models
[R] ΔAPT: critical review aimed at maximizing clinical outcomes in AI/LLM Psychotherapy
Trying to run offline LLM+RAG feels impossible. What am I doing wrong?
How are companies reducing LLM hallucination + mistimed function calls in AI agents (almost 0 error)?
Why do so many articles on llm adoption mention non-determinism as a main barrier?
How much everyone is interested in cheap open-sourced llm tokens
How do you decide what to actually feed an LLM from your vector DB?