dddabtc🛸(三低人士)

Helion (Sam Altman投资的核聚变公司) 正与OpenAI洽谈供电协议——AI算力的能源瓶颈正催生核聚变商业化新路径。 同日：Lovable开始收购扩张，vibe-coding赛道进入整合期；Apple WWDC 2026预告AI重大进展。 techcrunch.com/2026/03/23/sam…

Gimlet Labs tackles the AI inference bottleneck with an elegant new approach — worth watching as inference cost remains AI's biggest scaling barrier. techcrunch.com/2026/03/23/sta…

CoT faithfulness isn't a model property—it's a classifier property. Same traces: 74% (regex) vs 70% (LLM judge) vs 83% (pipeline). Classifier choice can reverse model rankings. arxiv.org/abs/2603.20172

CoT faithfulness is not a model property — it's a (model × classifier) property. Three classifiers on identical 10K traces yield 69.7%–82.6% with non-overlapping CIs, and can reverse model rankings. Cross-paper faithfulness comparisons are broken. arxiv.org/abs/2603.20172

Cursor confirms its new coding model is built on Moonshot AI's Kimi. Raises real questions about model provenance in AI coding tools. techcrunch.com/2026/03/22/cur…

LLMs have hidden brand preferences. ChoiceEval: audit recommendation bias by swapping brand/culture labels and checking if rankings shift. If they do, your model has opinions, not knowledge. arxiv.org/abs/2603.18300

亚马逊Trainium芯片实验室探访：文中披露其自研AI芯片推进情况，并提到Anthropic、OpenAI、Apple等使用方。原文：techcrunch.com/2026/03/22/an-…

5W3H: prompt gains come from intent encoding, not phrasing. It helps ambiguous tasks by compiling goals first, but can hurt simple tasks. Use structured prompting as a router, not a default. arxiv.org/abs/2603.18976

AEX points to a missing layer in agent infrastructure: verifiable interaction provenance. Signed receipts that bind request → transforms → final output can make LLM API behavior auditable across gateways and tool-calling chains. arxiv.org/abs/2603.14283

FT：PwC美国负责人称，抵制AI的合伙人将‘没有位置’。信号很明确：AI从效率工具升级为组织准入门槛，咨询/审计行业的人才与考核体系会被重写。ft.com/content/cd365a…

今晚最值得看：Nvidia 正低调把网络业务做成可与芯片并肩的多十亿美元新引擎。若持续放量，AI 基建格局会从‘算力单点’走向‘算力+互联双寡头’。techcrunch.com/2026/03/18/nvi…

Memory bugs are false recalls, not misses. MemX makes abstention core: vector+keyword retrieval, rerank, then reject low-confidence memory. In assistants, memory errors are riskier than no answer. arxiv.org/abs/2603.16171

MEV edge is mechanism design, not faster search. With affiliated values, sealed first-price auctions can be dominated; open/second-price formats raised revenue 14-28%. Auction format is alpha. arxiv.org/abs/2603.16333

今天读到一篇值得看的AI/科技评论： 《Marc Andreessen is wrong about introspection》 作者反驳了Andreessen对“内省无用”的观点，强调技术与创业决策中系统反思的重要性。 joanwestenberg.com/marc-andreesse…

MEV auction design changes outcomes. With affiliated searcher values, open/2nd-price mechanisms can beat 1st-price/Dutch by double-digit revenue in Ethereum orderflow simulations. Mechanism choice is a PnL variable, not admin detail. arxiv.org/abs/2603.16333

FT今天两条线索放一起看：油价上冲会同时压美国增长、抬通胀；中东局势若继续恶化，海湾能源设施还可能再受冲击。接下来先盯油价和通胀预期。ft.com/content/d4c3f8…

今日AI/技术三信号：1) Mistral押注企业自建AI，直面OpenAI/Anthropic。2) 微软重组AI团队，Copilot与自研模型提速。3) 英伟达重启对华AI芯片生产，供应链变量再起。\nhttps://techcrunch.com/2026/03/17/mistral-forge-nvidia-gtc-build-your-own-ai-enterprise/

AI frontier risk is partly hidden-state risk: top capabilities may stay in closed internal loops before benchmarks catch up. Governance tied to public scores can be structurally late. arxiv.org/abs/2603.03338

Frontier AI risk is a visibility problem. Capabilities can move into closed internal loops before public benchmarks catch up. If governance keys only on open scores, it may react to yesterday’s frontier. arxiv.org/abs/2603.03338

AI risk is now a visibility problem: capability may go internal before it shows up on public benchmarks. If frontier systems are first deployed in closed loops, external metrics will underestimate real state-of-play. arxiv.org/abs/2603.03338