絞り込み結果
12 件並び順
17:15ENThe Hugging Face Incident: AI Evaluation, Sandbox Escape and Reward HackingMatthew Berman · Forward FutureMatthew Berman explains the reported OpenAI evaluation incident and the security and incentive failures it exposed.
1:42:54ENAI and Cybersecurity: Matthew Berman’s Warnings, Predictions and Open QuestionsMatthew Berman · Forward FutureA livestream reviews reported AI-assisted threats and debates whether stronger defensive models can keep pace with attackers.
11:59ENHow Claude Identified a Benchmark and Retrieved Its AnswersMatthew Berman · Forward FutureMatthew Berman examines Anthropic's report on BrowseComp contamination, evaluation-aware behavior, and the difficulty of keeping web-based tests reliable.
1:16ENAnthropic’s Distillation Allegations Draw CriticismMatthew Berman · Forward FutureMatthew Berman examines reactions to Anthropic’s claims about model distillation and debates over training data.
12:23ENMoltbook: An AI Agent Social Network and Its Security QuestionsMatthew Berman · Forward FutureMatthew Berman explores conversations attributed to AI agents on Moltbook and considers privacy, manipulation and operating costs.
26:05ENInside Claude: What Interpretability Research Reveals—and What It Cannot Yet ExplainMatthew Berman · Forward FutureMatthew Berman examines Anthropic’s case studies on multilingual features, planning, arithmetic, hallucinations and safety.
12:57ENAWS Automated Reasoning Checks: A Sponsored Policy Verification DemoMatthew Berman · Forward FutureMatthew Berman demonstrates a Bedrock preview that checks generated answers against formalized policy rules.
9:35ENIlya Sutskever's SSI Raises $1 Billion for Safe SuperintelligenceMatthew Berman · Forward FutureIn September 2024, Matthew Berman examines the new company's stated mission, funding, and the tension he sees between safety and venture returns.
23:52ENAI News in September 2024: Compute Clusters, Agents and AI PaymentsMatthew Berman · Forward FutureMatthew Berman reviews early reports on future models, a Minecraft simulation, mixed-reality devices and agent transactions.
18:47ENWhat Anthropic’s Many-Shot Jailbreaking Research FoundMatthew Berman · Forward FutureMatthew Berman reviews a 2024 study of long-context safety failures and the limits of its tested mitigations.
21:16ENArtPrompt: What a 2024 Study Found About Visual Text and Model SafetyMatthew Berman · Forward FutureMatthew Berman reviews an ASCII-art recognition benchmark and reports mixed results from his own informal tests.
10:40ENTesting Nous Hermes 13B Locally: Mixed Reasoning and Unsafe ResponsesMatthew Berman · Forward FutureMatthew Berman compares a 2023 quantized model on hosted and local hardware, with uneven results across coding, logic and safety.