xAI quietly becomes a neocloud + curl maintainer pushes back on Mythos

TechCrunch's Sean O'Kane published a pointed editorial on the May 7 Anthropic-SpaceX deal, framing it as evidence that xAI is abandoning frontier model training to become a neocloud — renting GPUs to a rival rather than building competitive AI itself. The deal hands Anthropic all compute at xAI's Memphis-based Colossus 1 (220,000 NVIDIA GPUs, 300 megawatts) and is projected to generate $3 to $6 billion in annual revenue for the merged SpaceXAI entity. O'Kane reads it as "a major heat check before the IPO" and notes that Grok has limited traction outside X and is not used for enterprise tasks even internally at xAI. The interpretation matters: if xAI is no longer a serious frontier-model contender, the US frontier-lab landscape collapses from four labs to three (OpenAI, Anthropic, Google DeepMind).

Adam Dunkels — the engineer behind uIP, lwIP, and the Contiki OS — published a proof-of-concept in which Claude reads raw IP packet bytes and replies with valid ICMP ping responses, effectively acting as a user-space TCP/IP stack. The post documents Claude correctly parsing IP version, header length, TTL, protocol, and source and destination addresses, then constructing a properly formed reply packet that real ping clients accept. Latency is unsurprisingly enormous (this is a token-level interpreter, not a kernel module), but as a demonstration of what current frontier models can do with low-level byte manipulation, it's a clean data point on how far token-level reasoning has gotten — and a reminder that "what can a language model do" keeps pulling further from the obvious "language" framing.

Daniel Stenberg, the longtime curl maintainer, published a first-party assessment of Anthropic's Claude Mythos Preview after Project Glasswing scanned 178,000 lines of curl source code. The model surfaced 5 suspected vulnerabilities — reducing on review to 1 confirmed low-severity CVE plus roughly 20 bugs that were not vulnerabilities — and Stenberg concluded the bigger narrative around Mythos so far "was primarily marketing." His benchmark is direct: AISLE, Zeropath, and OpenAI's Codex Security have together driven 200 to 300 bug fixes for curl over the past 8 to 10 months, and he sees "no evidence that this setup finds issues to any particular higher or more advanced degree" than those tools. He does grant that all modern AI code analyzers — Mythos included — are substantially better than traditional static analyzers at finding security flaws.

In this week's Stratechery analysis, Ben Thompson argues that AI compute is bifurcating into three distinct workload categories that need fundamentally different hardware. Training keeps high-bandwidth GPUs (NVIDIA's lock-in); answer inference rewards token speed (Cerebras's WSE-3 packs 44 gigabytes of on-chip SRAM at 21 petabytes per second of bandwidth versus NVIDIA H100's 80 gigabytes of HBM at 3.35 terabytes per second); agentic inference, where humans aren't in the loop, mostly cares about memory capacity and cost-per-token at scale. Thompson treats Cerebras's revised IPO pricing of $150 to $160 per share (up from $115 to $125) and Anthropic's lease of 220,000 NVIDIA GPUs at SpaceX's Colossus 1 (the 300-megawatt Memphis data center) as concrete evidence that buyers are now sorting their compute spend by workload category rather than picking a single vendor.