Sam Altman on GPT-6 6️⃣, DeepSeek v3.1 🐋, Meta AI shake up 🧑‍💻

Sam Altman said that GPT-6 will arrive more quickly than the two-year gap between GPT-4 and GPT-5, emphasizing memory as the breakthrough advancement ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌  ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ 

Sign Up |Advertise|View Online TLDR Together With TLDR AI 2025-08-20 Your buyers have complex questions your website can finally answer (Sponsor) Today's SaaS buyers use AI everyday to answer questions and have no patience for a scavenger hunt when they visit your company's website. Concierge is a Perplexity-style answer engine that's trained on your company and lives on your website. It answers ultra-specific questions with accurate, personalized responses. Concierge handles any question, no matter how technical, with advanced RAG on your sources & media - in your brand language. Have control and visibility over every conversation, with guardrails and sentiment analysis. Modern brands use Concierge to build trust with buyers from the first moment. Turn every question into a conversation - and every conversation into revenue. 🚀 Headlines & Launches Sam Altman on GPT-6: 'People want memory' (3 minute read) Sam Altman said that GPT-6 will arrive more quickly than the two-year gap between GPT-4 and GPT-5, emphasizing memory as the breakthrough advancement. Personalization, specifically adjusting a model's political beliefs, may satisfy a new executive order that requires federal AI systems to be ideologically neutral. DeepSeek V3.1 just dropped — and it might be the most powerful open AI yet (12 minute read) DeepSeek quietly released DeepSeek V3.1 on Tuesday. The 685-billion parameter system challenges the dominance of American AI giants and reshapes the competitive landscape. Early performance tests revealed benchmark scores that rival proprietary systems from OpenAI and Anthropic. The model's hybrid architecture seamlessly integrates chat, reasoning, and coding functions into a single, coherent model. Mark Zuckerberg Shakes Up Meta's A.I. Efforts, Again (1 minute read) Mark Zuckerberg's restructuring creates separate teams for research, superintelligence, products, and infrastructure as Meta abandons its previous "Behemoth" frontier model to start fresh under new chief AI officer Alexandr Wang. The shake-up includes potential downsizing of the thousands-strong AI division and a strategic shift—Meta is now exploring third-party AI models after years of relying exclusively on its own technology. xAI readies Grok web with Imagine tool and Team profiles (1 minute read) Grok Imagine will enable users to generate images and short videos on the web. The feature, already accessible to mobile users, offers a dedicated gallery where people can view generations created by the model. xAI is close to releasing Team accounts, which will enable organizations to manage workspaces with isolated namespaces, dedicated chat histories, and collaborative projects. It fits into xAI's broader strategy of supporting both individual and enterprise workflows. 🧠 Deep Dives & Analysis Marketplace: my first attempt at training without backprop on GPU efficiently (41 minute read) People from a decade ago would hardly believe that we have an abundance of supercomputers in our homes. It's now possible to conduct end-to-end experiments in a solo project with modern hardware. We're entering into a new era of personal supercomputing. This article looks at an approach to training without backpropagation on GPUs efficiently - an experiment like this used to cost researchers an insane amount of money and resources just to test the idea. Faster MoE Training with Custom CUDA Kernels (20 minute read) Cursor rebuilt the Mixture-of-Experts layer from scratch using CUDA and PTX, achieving a 3.5x speedup in MoE operations. This resulted in a 1.5x end-to-end training speedup on Blackwell GPUs compared to Hopper. 🧑‍💻 Engineering & Research 1Password research finds that security leaders are struggling with unmanaged AI access (Sponsor) Our research found that only 21% of security leaders report having full visibility into the AI tools used in their organization. As organizations play catch-up with AI governance, they often find they're unequipped to enforce their policies. Read the research and learn how to mitigate AI access risks. Lemonade (GitHub Repo) Lemonade is a server that helps users run local LLMs with the highest performance by configuring state-of-the-art inference engines for their NPUs and GPUs. It supports both GGUF and ONNX models - Lemonade has a Model Manager that allows users to import custom models. Lemonade makes it easy to switch between configurations at runtime. It can be used with any OpenAI-compatible client library. Accelerating MoEs with Triton Grouped GEMM (6 minute read) PyTorch has introduced a cache-aware Triton BF16 Grouped GEMM kernel optimized for Mixture-of-Experts models like DeepSeekv3. The implementation offers up to 2.6× speedup over baseline PyTorch loops by batching independent GEMMs into a single kernel call. Signal and Noise: Reducing uncertainty in language model evaluation (7 minute read) Benchmarks with high signal (ability to distinguish between models) and low noise (consistency across training steps) are far more reliable for making scaling decisions, with some benchmarks showing 32% error reduction when filtered by signal-to-noise ratio. An analysis of 900K evaluation results across 465 models suggests that quality evaluation sets matter more than large sample sizes. Next Visual Granularity for Image Generation (2 minute read) NVG introduces a structured sequence-based framework for image generation, refining outputs progressively from global layout to fine details. 🎁 Miscellaneous Do LLMs Have Good Music Taste? (5 minute read) Claude models favor classic artists, especially jazz musicians like Herbie Hancock and Nina Simone. Reasoning models from OpenAI, xAI, and DeepSeek exhibit a bizarre preference for artists with numbers or dollar signs in their names, suggesting overly aggressive reinforcement learning may be creating unintended biases. Databricks says it's valued at over $100 billion in latest funding round (2 minute read) Databricks, now valued at over $100 billion, joins an exclusive club of private companies at this valuation alongside SpaceX and OpenAI. CEO Ali Ghodsi announced a new funding round exceeding $1 billion, with the company projecting $3.7 billion in annual revenue. This funding will support further AI product development, positioning Databricks against rivals like Snowflake and major cloud providers. ⚡ Quick Links Want more news from TLDR? (Sponsor) You'll probably like our flagship newsletter. It's all about tech, science, and programming. Same quick format. Still free. Subscribe now. Nvidia working on new AI chip for China that outperforms the H20 (3 minute read) Trump's recent openness to allowing more advanced chip sales creates a narrow window for regulatory approval. Top AWS chip designer reportedly defects to Arm as it weighs push into silicon (3 minute read) Rami Sinno ran engineering teams at Arm before joining Amazon - Sinno will now return to Arm. OpenAI CEO Sam Altman says that export controls alone won't hold back China's AI ambitions — "My instinct is that doesn't work" (2 minute read) Sam Altman says the US could be underestimating China's progress and capability in artificial intelligence. Introducing Chat Mode (1 minute read) You can now build text-only conversational agents on ElevenLabs. Love TLDR? Tell your friends and get rewards! Share your referral link below with friends to get free TLDR swag! https://refer.tldr.tech/0b6a6dc1/2 Track your referrals here. Want to advertise in TLDR? 📰 If your company is interested in reaching an audience of AI professionals and decision makers, you may want to advertise with us. Want to work at TLDR? 💼 Apply here or send a friend's resume to jobs@tldr.tech and get $1k if we hire them! If you have any comments or feedback, just respond to this email! Thanks for reading, Andrew Tan, Ali Aminian, Jacob Turner & Sahil Khoja Manage your subscriptions to our other newsletters on tech, startups, and programming. Or if TLDR AI isn't for you, please unsubscribe.