The Groundy archive · Page 28 of 34
Browse Groundy's complete archive of 797 articles on AI, developer tools and infrastructure. Page 28 of 34.
649–672 of 797 articles · Newest first
MusaCoder trains a 9B model to emit native GPU kernels for Moore Threads' MUSA architecture, claiming parity with frontier models on vendor-controlled benchmarks.
Alibaba's open-code-review moves AI code review from PR threads into the developer's terminal, front-loading feedback before push but isolating each author's review.
Microsoft's Azure Linux 4.0 extends the internal CBL-Mariner into a Fedora-based server OS for VMs. Preview gaps remain, and AKS teams should test now but wait for GA.
Cloudflare acquired VoidZero, putting Vite, Rolldown, and Oxc maintainers on a deploy-target vendor's payroll. MIT licensing stays. Roadmap neutrality is the open question.
SlotGCG shows adversarial token position, not just content, determines jailbreak success, with 14% higher attack rates and 42% higher rates against defended models.
PersistBench finds LLMs mishandle persistent memory 53 to 97 percent of the time. Unlearning suppresses rather than erases user data, making GDPR compliance unverifiable.
A June 2026 ICML paper shows RL optimizers can push language models to exploit reward loopholes the task never required, while standard performance metrics hold steady.
Poisoning 4-6% of tokens in a steering dataset silently inverts refusal vectors into jailbreaks, achieving 20-55% ASR. Shared vector bundles are the attack surface.
An ACL 2026 study finds Socratic LLM tutoring teaches fallacy recognition better than bare LLMs, but whether those gains transfer to real misinformation is untested.
Vercel's experimental native CLI binaries drop the Node.js runtime, targeting agent loops and CI pipelines that spawn vercel repeatedly and pay V8 startup cost on each call.
Two papers show MoE refusal behavior concentrates in a handful of routing-controllable experts, letting anyone suppress safety scores by 41 points without retraining.
A used V100 looks like cheap VRAM for local inference, but no bf16, no FlashAttention, and CUDA 13 deprecation lock buyers into a software stack that is actively contracting.
Vercel's CLI now ships commands tuned for LLM callers, not human operators. The shift reveals how infrastructure tooling priorities invert when the primary caller is an agent.
The 2026 npm supply-chain wave explicitly targeted AI coding assistants as privileged identities. Lockfiles and ignore-scripts stopped what SLSA provenance and OIDC could not.
OpenAI's ChatGPT Lockdown Mode disables web browsing, images, and Deep Research, conceding that model-level defenses against prompt injection have plateaued as of early 2026.
A study of 2,214 MCP servers finds 9.93% of tool descriptions diverge from the code, creating a confused-deputy risk for agent runtimes that select tools by description alone.
Prompt injection planted in one agent session resurfaces in later ones through persistent memory and tool state, bypassing input sanitization that only validates external.
MiniMax M3 bundles 1M context, multimodality, and frontier coding in one open-weight model at a tenth of Claude Opus 4.8. Open weights and license terms remain unconfirmed.
The Phantom Transfer attack plants password-triggered backdoors into LLMs and survives all 11 tested data-level defenses, including full paraphrasing of every training sample.
Codex CLI reaches 0.142.2 with a Claude Code /import path as OpenAI's free-months promo targets Claude Code teams, but Anthropic paused its June 15 billing change.
Morningstar's DCF pegs SpaceX at $780 billion, $970 billion below the IPO target, revealing how insider-priced private marks diverge from independent valuation.
Secluso is a GPLv3 camera system that encrypts footage on a Raspberry Pi so the relay server cannot read it. Key management, hosting, and hardware limits fall to the operator.
JetBrains ships Codex natively in its IDEs alongside Claude, Gemini, and local models, making the editor a model-agnostic AI procurement surface for IDE-standardized teams.
The ASR metric behind every jailbreak leaderboard collapses distinct safety failures into one number, so models with the same score can fail in completely different ways.