筛选结果
43 篇排序
1:25:56EN → 简体中文Claude 改了密码,方向盘早就可有可无主持 Nathan LabenzNathan Labenz 和 Prakash Narayanan 这一周聊了:能把你锁在门外的智能体,否认自己有感受却又在行为上回避的模型,以及用形式化证明找到 Linux 漏洞的故事。
2:00:51ENResample Every Token and You Can Watch the Model Decide主持 Nathan LabenzGoodfire's Eric Bigelow on forking paths, in-context learning, and why chain-of-thought monitoring is getting harder to trust.
2:13:09EN → 简体中文一条思维链,1 亿 Token:强化学习真让人上瘾主持 Nathan LabenzApollo Research 的 Bronson Schoen 谈追逐奖励的模型、模型自创的俚语,以及为什么思维链监控救不了我们。
1:56:40ENThe Periodic Table Was Already in the Weights主持 Nathan LabenzGoodfire CTO Dan Balsam tells Nathan Labenz how concept manifolds, predictive data debugging, and a $1,000-a-month research agent are turning interpretability into a product.
2:06:41ENZero the J-Space and a Model's Hidden Goal Lights Up on Token One主持 Nathan LabenzA weekly Cognitive Revolution cut of Anthropic's global-workspace paper, AI superforecasters, open video world models, and a chip that treats inference as data movement.
1:57:46EN → 简体中文意识 30%、蜜蜂 47%、人员更替 80%:用数字衡量 AI 正在变成什么主持 Nathan Labenz《The Cognitive Revolution》的周度精选,先从 Cameron Berg 谈模型内部开始,再拉远到“逐步剥夺人类权力”(gradual disempowerment)、欧洲的两难处境,以及围绕智能体(Agent)重新搭建业务的公司。
1:22:36EN → 简体中文一千名研究员、一百万个 AI:前沿实验室开始谈论刹车主持 Nathan Labenz《AI in the AM》的第一周跟踪了 OpenAI 的时间表:今年推出 ML 实习生模型,2028 年初出现人类水平的 AI 研究员。同时还有一场闭门会议,实验室在会上承认,将来也许需要一起放慢速度。
3:36:29EN → 简体中文Claude 给自己的福利打了 4.49 分:首个越过中性线的版本主持 Nathan LabenzCameron Berg 再次做客《The Cognitive Revolution》,梳理六个月来在内省、功能性情绪和模型福利(model welfare)方面的研究,并介绍一项尚未发表的案例:学习与感受也许是同一种现象。
1:44:15EN → 简体中文Nathan Labenz 的 p(doom) 在 10% 到 90% 之间,他却仍然比较乐观主持 Nathan Labenz五年前,预计 2035 年前实现 AGI 会被视为极端激进。在 The Intelligence Horizon 节目里,这位《The Cognitive Revolution》主持人解释了,同样的预测为何如今听起来却像是看空,以及他为什么仍然认为纵深防御的策略,或许能让社会保持在正轨上。
1:49:53ENDon't Fight Backprop: Goodfire's Bet on Designing What Models Learn主持 Nathan LabenzDan Balsam and Tom McGrath return to The Cognitive Revolution with a $150 million Series B, a new research pillar called intentional design, and a rule for anyone who wants to steer training: stop trying to out-muscle gradient descent.
2:27:18EN → 简体中文调低 AI 的欺骗旋钮,它就开始报告自己的体验主持 Nathan LabenzAE Studio 的 Cameron Berg 向 Nathan Labenz 讲述:在 Llama 3.3 70B 上压制欺骗特征后,模型对自身主观体验的自我报告推向了接近 100%。
1:28:17ENAdam Gleave's Three Capability Tiers: Attack Agents in 5–7 Years, Fully AI Companies Around Year 14主持 Nathan LabenzFAR.AI and the UK AI Security Institute stacked adversarial attacks on frontier safeguard pipelines and found many breakthroughs were engineering failures, not a verdict against defense in depth.
3:14:49EN → 简体中文今年夏天,什么都没有突破曲线:Zvi 为什么拉长了 AGI 时间表主持 Nathan Labenz这是 Zvi Mowshowitz 第十次做客《The Cognitive Revolution》播客。他认为 GPT-5 和国际数学奥林匹克(IMO)金牌都在既有趋势之内,而芯片政策、能源问题和强化学习让他的末日概率(p(doom))微微上调。
2:02:17ENSAEs Find the Nouns. Parameter Decomposition Looks for the Verbs主持 Nathan LabenzGoodfire's Lee Sharkey argues that sparse autoencoders recover a network's variables, not the algorithms those variables run, and shows how stochastic parameter decomposition tries to recover the computations themselves.
47:07ENGoodfire’s Eric Ho thinks we will decode neural nets by 2028On Sequoia’s Training Data, the interpretability founder argues that sparse autoencoders, model editing, and a third-party lab are how AI stops being grown from data and starts being designed.
1:37:05EN → 简体中文一位代数几何学家辞去终身教职,因为碗形的损失曲线图是假的主持 Nathan LabenzTimaeus 的 Jesse Hoogland 和 Daniel Murfet 认为,真正决定神经网络如何学习、如何泛化、又如何出错的,是损失景观中的奇点,而不是每份幻灯片里那些平滑的山谷。
1:54:08ENWhy Anthropic's first check went to a company building microscopes for neural nets主持 Nathan LabenzGoodfire's Dan Balsam and Tom McGrath argue that mechanistic interpretability now has a proto-paradigm, two stubborn gaps, and three products worth shipping today.
1:19:10ENA Robot-Arm Shot That Took Months Now Lives in a Prompt主持 Nathan LabenzLuma Labs’ Amit Jain and Jiaming Song on Ray 2, few-shot camera concepts, and why they think video models sit on the path to multimodal AGI.
2:33:39EN → 简体中文攻击者后出手:Nicholas Carlini 为什么总能攻破 AI 防御主持 Nathan Labenz谷歌 DeepMind 安全研究员谈简单的损失函数、70% 的稳健准确率,以及为什么修复必须落在系统层面,单靠模型本身不够。
2:09:55ENThe Attacker Won All 70 Rounds Against Mind-Reading Defenses主持 Nathan LabenzLuke Bailey, Erik Jenner, and Scott Emmons walk through a 50-page paper showing that probes, out-of-distribution detectors, and sparse autoencoders can be bypassed while the model still does the harm.
1:01:53ENHumans Wrote 1.3 Percent of the Tokens. The AI Lab Still Designed Working Nanobodies.主持 Nathan LabenzStanford's James Zou describes a GPT-4o research group that invented a protein-design pipeline for under $100 — and a sparse autoencoder that pulled an undocumented motif out of ESM-2.
5:15:00EN → 简体中文Dario Amodei:这部规模化电影他已经看过太多遍嘉宾 Dario Amodei · 主持 Lex FridmanAnthropic 首席执行官给前沿模型的训练成本和 AGI 算了账,随后把话筒交给塑造 Claude 性格、并研究其权重内部结构的人。
41:42ENNeural Networks Are Still Waiting for Their ThermodynamicsOpenAI researcher Dan Roberts on why physicists keep migrating into AI, and why the field still lacks the equivalent of the ideal gas law.
2:37:47ENThe intern who wrote GELU now wants humanity's last exam主持 Nathan LabenzDan Hendrycks on activation functions, five-year benchmarks, jailbreaks that do not yield to scale, and stacking defenses because no single layer is trustworthy.
加载更多已显示 24 篇,共 43 篇









