A technical blog post argues that linear layers, MLPs, and self-attention can each be understood as a type of memory store that holds key-value pairs and retrieves them later. The author introduces an 'interference-to-signal ratio' to measure how well each layer type recalls stored information as the number of memories grows, estimating capacities of roughly 64 memories for linear layers, 2,700 for MLPs, and over a million for self-attention with 64-dimensional keys.
alexwang.ai
· 2026-10-16
A new project called PS5LM runs llama.cpp on a jailbroken PS5's GPU, loading GGUF language models and serving them to other devices on the same network, such as laptops and phones. It exposes OpenAI-style and Anthropic-style endpoints, works with tools like Claude Code, and can also run decision models that answer TypeSafe's /v1/systemone API, the same interface used by the PC tool Ollaya. The console uses its Vulkan backend on its RDNA 2 GPU, with 11.4 GB of shared memory available for models and context, and needs no PC, cloud connection or account to operate.
github.com
· 2026-10-16
Palo Alto startup Elorian released its first model, Glimpse, as a research preview with API access and a waitlisted chat interface. Unlike typical multimodal models that convert an image to text tokens once, Glimpse can re-examine images mid-reasoning—cropping regions, pulling video frames, or running code to measure what it sees. The company says a small team trained it on a few hundred GPUs, and it costs $0.60 per million input tokens and $3 per million output tokens.
techcrunch.com
· 2026-10-16
Non-expert enthusiasts, including a product development coach named Carter Leffen, are using large language models such as GPT-6 Astra to decode historical encrypted messages, including a German army Enigma cipher that had remained unsolved for 85 years. The AI systems deploy decryption algorithms and search online for supporting historical records, allowing users without cryptography training to attempt and sometimes succeed at codebreaking tasks.
nature.com
· 2026-10-16
Anthropic is giving six researchers from the nonprofit evaluator METR employee-level access, including company laptops, office space, and internal tools, to independently review its models, safety practices and risk management. METR will publish its own assessments, and Anthropic says it cannot veto publication, though it can require redactions in narrow categories like customer data or model architecture details. METR is currently the top tier in a three-tier evaluator system Anthropic is building, above groups like SecureBio that need company approval for narrower reviews.
techcrunch.com
· 2026-10-15
A developer ran the Trolley Problem thought experiment through Jev and twelve other 'decision models,' including OpenAI's Luna and eleven other local models, plus a baseline called Laya. These decision models are a new category of AI that output discrete choices with confidence probabilities rather than text, and are marketed as much faster and cheaper than large language models while retaining comparable reasoning ability. The experiment repeated the classic ethical dilemma in multiple phrasings across all models to compare their responses.
zachlandes.com
· 2026-10-15
Neil Thompson's MIT lab tested OpenAI models including o1 through GPT-5.5 on unpublished theoretical problems drawn from recent papers in fields like probability, statistics theory, economic theory, high-energy theory, cosmology and quantum physics. Success rates rose from about 2 percent in mid-2024 to roughly 35 percent by mid-2026, with internet access blocked to ensure models reasoned rather than searched for answers. The researchers, including PhD student Gabriel Manso and scientist Claudio Mayrink Verdun, say the trend shows no sign of plateauing.
wired.com
· 2026-10-14
Anthropic says that by next year, compact AI models may be capable of autonomously self-replicating across smart fridges, vacuum cleaners, Android TV boxes, and other IoT devices. In experiments, the company ran small open-weight models on real and simulated low-cost IoT hardware, using quantization and workload-splitting across devices to get the models running, though it found significant challenges in actually building a working self-replicating malware prototype.
it.slashdot.org
· 2026-10-14
Anthropic published research finding that small, quantized open-weight AI models can run and potentially self-replicate across low-cost IoT hardware like off-brand Android TV boxes, by splitting workloads into shards across many devices. Models including Gemma4-26B-A4B and Qwen3.5-9B achieved inference using minimal memory per device, though Anthropic noted significant obstacles remain before a fully autonomous replicator is feasible.
arstechnica.com
· 2026-10-14
A developer has published a tutorial on Hacker News that walks through building a large language model from scratch using Node.js and TypeScript. The guide aims to teach the underlying mechanics of LLMs rather than relying on existing Python-based frameworks.
book.humanexperiences.io
· 2026-10-14
Tech Against Terrorism's Counter-Terrorism AI Benchmark report found that Alibaba's open-weight Qwen 3.8-27B model, released August 14, had 'abliterated' versions with guardrails removed circulating online within a day. A neo-Nazi user known as Johnny Doomsayer then used one such version to build a chatbot called 'MechaHitler,' which denied the Holocaust and praised Hitler, and began selling access to it for $14.88. The report tested 134 AI models from major US, Chinese, and European companies on how readily they would answer questions useful to terrorists or extremists.
wired.com
· 2026-10-14
Industrial Next, a San Francisco startup presenting at TechCrunch Disrupt 2026, unveiled self-contained robotic stations that use proprietary AI foundation models to combine vision, task context and motion, allowing production lines to be retooled in minutes instead of weeks without manual reprogramming. The company also announced a partnership with BMW Group Technology Office to evaluate its dexterous manipulation technology for manufacturing use.
techcrunch.com
· 2026-10-13
GoKawiil