Loading...
Loading...
An undergraduate computer scientist expresses growing disillusionment with recent mechanistic interpretability (mech interp) research, particularly work from Anthropic, despite initial enthusiasm for techniques like sparse autoencoders and attribution graphs. The poster questions the field’s current direction, suggesting some high-profile papers may overclaim or lack practical rigor, and wonders whether mech interp is maturing into reliable engineering knowledge or remaining exploratory. This ma
Mechanistic interpretability aims to make large models transparent and trustworthy, which affects model safety, debugging, and deployment decisions. Tech professionals need to judge whether new methods yield reliable engineering knowledge or remain exploratory before adopting them.
Dossier last updated: 2026-05-16 01:07:47
NVIDIA disclosed its Q1 13F holdings as of March 31, 2026, revealing major portfolio moves that signal strategic bets on AI infrastructure. The largest change: a 94.5% increase in CoreWeave (CRWV) to 47.2 million shares, reflecting confidence in rising AI cloud GPU demand and CoreWeave’s role as a top NVIDIA GPU renter. NVIDIA also newly disclosed a 7.8 million‑share position in optical module maker Coherent (COHR), aligning with surging data‑center interconnect needs. Other sizeable holdings include Intel (INTC) at 214.8 million shares, Nokia (NOK) at 166.4 million, Synopsys (SNPS) at 4.8 million, Nebius (NBIS) at 1.2 million, and Generate Biosciences (GENB) at 833k. The moves underscore NVIDIA’s investments along its hardware and data‑center ecosystem.
An undergraduate computer scientist expresses growing disillusionment with recent mechanistic interpretability (mech interp) research, particularly work from Anthropic, despite initial enthusiasm for techniques like sparse autoencoders and attribution graphs. The poster questions the field’s current direction, suggesting some high-profile papers may overclaim or lack practical rigor, and wonders whether mech interp is maturing into reliable engineering knowledge or remaining exploratory. This matters because mechanistic interpretability aims to reveal how large models work—insights that affect safety, debugging, and regulatory oversight—and skepticism from emerging researchers could influence community norms, research priorities, and recruitment. The post invites discussion on standards, reproducibility, and the balance between theory and usable tools.
CoreWeave reported record Q1 FY2026 revenue of $2.078 billion, up 111.6% year-over-year, but posted a net loss of $740 million, a 134.9% widening from the prior year. The company saw adjusted EBITDA of $1.157 billion (56% margin) and operating cash flow of $2.984 billion, while backlog reached $99.4 billion and contracted power exceeded 3.5 GW with a 1 GW effective deployment. Operating expenses surged—driven by technology and infrastructure—and net interest expense doubled to $536 million. CoreWeave also closed major strategic deals (including a $21 billion agreement with Meta and multi-year work with Anthropic), raised an $8.5 billion non-recourse term loan and received a $2 billion equity investment from Nvidia to fund aggressive power and capacity expansion through 2030.
Anthropic has launched Claude for Microsoft 365, bringing Claude integrations to Word, Excel, PowerPoint and a beta for Outlook across Windows and Mac for paid plans. The assistant can follow users across open Office documents, persist conversations tied to each file, and transfer changes among a spreadsheet, memo and presentation without re-explaining tasks. In Outlook, Claude can triage mail into items that need user reply, can be drafted on the user’s behalf, or are spam; suggested replies appear as drafts that the user must review before sending. Administrators can deploy the add-ins via Microsoft AppSource and Microsoft Admin Center. This tight Office integration aims to streamline cross-app workflows and accelerate document-centric productivity.