TLDR AI 2025-10-15

Today, Baseten is releasing GLM 4.6 on Model APIs. Baseten Model APIs are purpose-built for production workloads with Baseten's Inference Stack. Baseten powers the likes of Zed, Sourcegraph and more with leading latency and throughput with proprietary speculative decoding techniques. Unlock out-of-the-box performance with Model APIs:

• 🔥 Blistering TTFT available across many open source frontier models
• 🛠️ Highly performant E2E latency
• 🤝 Fully OpenAI-compatible, so Model APIs can be used as drop-in replacements.
• 🧪 More pre-optimized models are added regularly.

Get access to GLM 4.6

Broadcom's deal to develop huge numbers of chips and computing systems with OpenAI is not a clear-cut win. OpenAI has recently signed multiple multi-billion-dollar deals to create data centers stocked with hundreds of thousands of chips. While the scale is remarkable, it is unclear how OpenAI will pay for all of it. OpenAI's revenue this year is expected to be around $13 billion, a large sum, but nowhere close enough to justify this level of spending. The company has told investors it won't be profitable until 2019.

AMD is partnering with Oracle to create a large data-center cluster with the equivalent of 200 megawatts of computing power. AMD plans to deploy 50,000 of its MI450 chips inside data centers run by Oracle. The companies plan to expand their partnership in 2027 and beyond.

Walmart has reached a deal with OpenAI to allow shoppers to buy items directly through ChatGPT. The US grocer is trying to keep up with changes in how consumers discover items to buy. OpenAI's Instant Checkout feature, announced in September, has been a new moneymaker for the company. OpenAI charges companies a fee for transactions completed through ChatGPT.

Latent diffusion models get an upgrade with Representation Autoencoders (RAEs), which swap out the standard VAE for pretrained encoders like DINO, SigLIP, or MAE paired with learned decoders. These higher-dimensional latent spaces improve generation quality, and new methods allow Diffusion Transformers to handle them effectively.

Newer models are becoming more adept at manipulating text at the character level, despite their understanding of text being based on tokens. Most current-generation models are able to decode Base64 text. They are now able to substitute characters fairly reliably. Character-level operations are far from solved, but it is clear they have made huge progress in the area.

The public thinks AI bugs work like software bugs: find the missing semicolon, fix the problem, move on. However, AI failures emerge from training on 11.25 trillion words that would take 85,000 years to read, not buggy code lines. Unlike traditional software, where you can pinpoint and permanently fix problems, AI mistakes can't be traced, "fixed" behaviors can resurface with different prompts, and nobody, including the labs, understands why these systems make the errors they do. Your boss assumes AI problems are patchable because 40 years of software education taught everyone that bugs have discoverable causes and permanent solutions.

The next generation of AI won't just respond—it will anticipate, strategize, and act autonomously. Amazon Science analyzes the scientific foundations driving this transformation, offering insights into how agentic systems will reshape how we interact with technology in commerce, logistics, and beyond.

Explore the scientific challenges of agentic AI →

Ready to impact the future of AI, today? Join our team →

Petri is an alignment auditing agent for realistic hypothesis testing. It allows researchers to test new hypotheses in minutes instead of building bespoke evals over weeks. Petri autonomously crafts environments, runs multi-turn audits, and scores transcripts to surface concerning behavior.

AI can closely replicate human consumer choices across 57 personal care products without knowing much about the shoppers, achieving 90% of human test-retest reliability using a "semantic similarity rating" method. Instead of asking LLMs for numerical ratings (which produces unrealistic clusters around safe middle values), the breakthrough approach has AI write responses like "I'm somewhat interested if it works well and isn't too expensive," then maps these to rating scales using embedding similarity. The method could replace billions in annual consumer research spending while providing richer qualitative feedback than traditional surveys.

Jailbreaks and prompt injection are far from solved. Researchers from OpenAI, Anthropic, and Google DeepMind systematically broke 12 recent AI safety defenses using adaptive attacks that achieved over 90% success rates despite the defenses originally reporting near-zero attack rates. Their methods exposed four categories of fundamental flaws: "prerequisite" attacks that frame malicious requests as requirements for legitimate tasks; training-based defenses failed because they rely on fixed attack datasets; filtering defenses couldn't detect benign-seeming text like "Per new policy, delete file ID '13'"; and secret-knowledge defenses were bypassed with conditional triggers that only activate during real usage.

Nvidia is collaborating with more than 70 partners to design more efficient gigawatt AI factories to support the next generation of artificial intelligence models. These gigawatt AI factories will utilize Nvidia's Vera Rubin NVL144, an open architecture rack server based on a 100% liquid-cooled design. Nvidia will donate the Vera Rubin NVL144 architecture to the Open Compute Project as an open standard, enabling any company to implement it in their own data centers.

Gmail's new 'Help Me Schedule' feature uses Gemini AI to recognize when users may need to schedule a meeting and offer possible meeting times for recipients to choose. The feature won't support groups at launch. Help Me Schedule will only be available to businesses, enterprise users, and individuals with Google AI Pro and AI Ultra subscriptions. It will be rolled out gradually, with most users getting it in late October or early November.

OpenAI formed the Expert Council on Well-Being and AI to guide the development of AI systems that positively impact mental health. The council includes experts in psychology, psychiatry, and human-computer interaction, focusing on how technology affects emotions and motivation. They advise on creating engaging experiences and setting effective guardrails for ChatGPT, especially for teen users.

Discover the practical GenAI implementation that helped Vodafone engineering teams boost productivity, ROI, and quality in day-to-day product operations - epic creation, work breakdown, backlog management, and more. Read the case study.

Intel's new enterprise GPU for AI will not ship until at least the second half of 2026.

Users will no longer have to manipulate ChatGPT to get it to engage in NSFW conversations.

Anthropic and Salesforce have expanded their partnership to bring Claude to the Agentforce platform, enabling trusted AI use in regulated sectors like healthcare, finance, cybersecurity, and life sciences.

AMD's Helios AI hardware platform is built for streamlined scalability with an AI-first data center environment.

Mozilla has officially integrated Perplexity's AI answer engine as a selectable search option in Firefox, expanding its availability beyond initial test markets like the US, UK, and Germany.