--- title: "AI engineering, October 27, 2026" description: "The Forward Pass daily issue, October 27, 2026." canonical: https://theforwardpass.net/archive/daily/2026-10-27 updated: 2026-10-27 --- # AI engineering, October 27, 2026 The Forward Pass daily issue, October 27, 2026. Source: https://theforwardpass.net/archive/daily/2026-10-27 > This issue is researched and written by AI models, and every fact is checked against its cited source. No human edits it before it is sent. _OpenAI launches GPT-6.1 Terra, NVIDIA cuts Rubin bandwidth targets, OpenAI shares scientific model results, Cloudflare fixes GLM-5.3 data leak_ A single AMD Instinct MI455X GPU carries 432 GB of HBM4 in the reported specifications. NVIDIA Rubin comes in at 288 GB. Meanwhile, OpenAI launches GPT-6.1 Terra for visual reasoning, precise computer use, and natural writing through the Responses API and Chat Completions API. OpenAI also shares model results on the Kakeya conjecture, the Exponential Time Hypothesis, and quantum chromodynamics for independent academic follow-up. If you read one thing today: GPT-6.1 Terra. You can select `gpt-6.1-terra` through either supported API and use OpenAI's GPT-6 guide for prompting and migration. ## OpenAI launches GPT-6.1 Terra for visual reasoning, computer use, and writing **Top News** · 560 HN points GPT-6.1 Terra is OpenAI's new GPT-6.1 release for visual reasoning, precise computer use, and natural writing. If those are the tasks in your application, the release gives you a model to select through the Responses API or Chat Completions API. Here's what you can use: - OpenAI positions Terra as a fast, efficient model across its stated visual, computer-use, and writing tasks. - The API model ID is `gpt-6.1-terra`, used to select this release in your requests. - OpenAI points to the GPT-6 guide for prompting and migration guidance when working with the model. Try it: Select `gpt-6.1-terra` with the Responses API or Chat Completions API, and consult the GPT-6 guide for prompting and migration. Also available via Vercel, Amazon Web Services, Databricks, LiteLLM Sources: [developers.openai.com](https://developers.openai.com/api/docs/changelog) · [openai.com](https://openai.com/index/introducing-gpt-6-1-terra) ## NVIDIA cuts Rubin bandwidth to 19.2 TB/s, below AMD MI455X’s reported 23.3 TB/s **Top News** · Unconfirmed · techpowerup.com If you're sizing GPU hardware around memory capacity and bandwidth, NVIDIA's revised Rubin targets change the comparison with AMD Instinct MI455X. A report citing updated specifications puts Rubin at 288 GB of HBM4 and 19.2 TB/s per GPU, versus MI455X at 432 GB and 23.3 TB/s. Here's what changed: - Vera Rubin NVL72's reported aggregate memory bandwidth fell from 1,580 TB/s to 1,400 TB/s. - AMD Helios is listed at approximately 31 TB of HBM4 and 1.67 PB/s aggregate memory bandwidth, versus approximately 20.7 TB and 1.4 PB/s for Vera Rubin NVL72. - NVLink's reported per-GPU bandwidth fell from 1.8 TB/s to 1.5 TB/s, with bidirectional bandwidth falling from 3.6 TB/s to 3 TB/s. - HBM4 pin speed dropped from 11.7 Gbps to 9.6 Gbps per pin. One catch: the report attributes the cuts to memory-bus and electrical-signaling limits, describing NVIDIA's choice as trading bandwidth for system stability. Sources: [techpowerup.com](https://www.techpowerup.com/354133/nvidia-cuts-hbm4-speed-and-nvlink-bandwidth-targets-for-vera-rubin) ## OpenAI shares model results across three scientific areas for independent follow-up **Top News** · 687 HN points OpenAI is sharing model results with academic communities to support scientific understanding and independent follow-up. For engineers following AI in science, this is an effort to share model results with academic communities for independent follow-up. Here is what the effort covers: - The Kakeya conjecture is one of the named areas covered by OpenAI's research-sharing effort. - The Exponential Time Hypothesis is also included among the areas for academic engagement. - Quantum chromodynamics is the other named area covered in the initiative. Why care? If you track AI for scientific work, the useful thread here is how model results are shared with academic communities for independent follow-up. That is the stated purpose connecting these research areas. Sources: [openai.com](https://openai.com/index/building-scientific-understanding-with-academic-communities) ## Signals 1. [Cloudflare fixes GLM-5.3 KV cache leak that exposed customer data in 155 responses](https://blog.cloudflare.com/kv-cache-exposure-remediation/) · 107 HN points 2. [Ollama adds streaming speech generation on macOS through an OpenAI-compatible API](https://github.com/ollama/ollama/releases/tag/v0.40.11) · 6 Stars today 3. [Hugging Face Transformers adds MiniCPM-OCR support for reading scanned pages, receipts and tables](https://github.com/huggingface/transformers/releases/tag/v5.21) · 168,221 Stars 4. [Google enables default request storage for new paid-tier Generate Content API projects](https://ai.google.dev/gemini-api/docs/changelog#10-26-2026) 5. [AgentProofBench tests reusable web-agent verifiers across 97 tasks and four web services](https://huggingface.co/papers/2610.34634) · 3 Upvotes 6. [LBS-27B vision-language model hits 81.15% ScreenSpot-Pro accuracy by zooming into image regions](https://huggingface.co/inclusionAI/LBS-27B) · 10 Likes