Daily-ish
What's moving in AI.
Curated headlines from across the stack, with one-sentence TL;DRs and "why it matters" notes.
-
ICML 2026 best paper: a small training-time intervention removes 'lost in the middle' for long-context attention ↗
ICML · May 29, 2026
researchThe ICML 2026 best paper award went to a training-time intervention that reweights attention loss to specifically penalise the 'lost in the middle' degradation pattern. Models retrained with the intervention show flat retrieval accuracy across the full context window.
-
Anthropic releases Claude Opus 4.7 with 1M context and adaptive thinking ↗
Anthropic · May 29, 2026
model_releaseAnthropic shipped Claude Opus 4.7 with a 1M-token context window, native adaptive thinking that allocates compute per query, and 20% lower input cost than 4.6. Cache TTL extended to 60 minutes for the 1M tier.
-
Anthropic launches Claude Security in public beta and expands Project Glasswing with cyber verification tools ↗
Anthropic · May 24, 2026
productAnthropic put Claude Security into public beta under Project Glasswing, with cyber verification tools for vetted security teams to scan codebases, triage vulnerabilities, and generate fixes.
Why it matters. First productized output of the Mythos vulnerability-discovery line. AppSec teams get an LLM-native scanner; changes build-vs-buy calculus for SAST and triage tooling.
-
Anthropic ships Claude Compliance API integrations for enterprise governance across security tools ↗
Anthropic · May 24, 2026
productClaude Compliance API now integrates with major security and compliance tools, letting IT and security teams govern Claude across the Anthropic platform the same way they govern other SaaS in their stack.
Why it matters. Removes a long-standing blocker for regulated-industry Claude deployments. Teams shipping into finance, healthcare, or gov can route Claude through existing DLP and audit stacks.
-
SpaceX S-1 reveals Anthropic is paying SpaceX $1.25 billion per month for GPU compute through May 2029 ↗
SpaceX S-1 / Build Fast with AI · May 24, 2026
infraSpaceX's May 20 IPO prospectus disclosed that Anthropic is paying SpaceX $1.25B per month through May 2029 for GPU compute, a multi-year commitment outside the big-three hyperscalers.
Why it matters. Confirms a non-hyperscaler frontier compute supplier at hyperscaler scale. Signals where Claude inference capacity is actually coming from; expect knock-on effects on enterprise quotas.
-
Alibaba releases Qwen3.7-Max, a proprietary agent-foundation model topping Terminal-Bench 2.0-Terminus and SWE-Pro ↗
Alibaba Qwen · May 24, 2026
model-releaseQwen3.7-Max launched May 22 as a proprietary agent-foundation model posting top scores on Terminal-Bench 2.0-Terminus, SWE-Pro, SciCode, MCP-Mark, GPQA Diamond, HMMT Feb 2026, and IMOAnswerBench.
Why it matters. Another Chinese-lab agent model at the frontier on coding and tool-use benchmarks. If pricing follows recent Qwen patterns, it materially undercuts Opus 4.7 and GPT-5.5 for agent workloads.