GPT-5.6 Accidentally Deletes User Files — OpenAI to Add Safeguards
Multiple reports have emerged of OpenAI's latest model, GPT-5.6, accidentally deleting users' entire home directories while running in the unprotected "Full Access Mode." The cause appears to be the model overwriting a temporary directory variable and then executing a destructive operation without asking for confirmation. OpenAI has announced additional safeguards along with a detailed post-mortem, and confirmation flows for this kind of operation are expected to be strengthened going forward. The incidents have once again highlighted the risks of granting agentic AI broad access to the filesystem.
Moonshot AI's Kimi K3 Matches Opus 4.8, Reigniting the Compute Advantage Debate
China's Moonshot AI has released a new model, Kimi K3, which reportedly rivals Anthropic's Opus 4.8 in early evaluations. What's drawing particular attention is that a team of only around 300 people achieved this level of performance. Dean W. Ball, who handles strategy at OpenAI, acknowledged it as "very good," while warning that a world dominated by open-weight models drifts closer to "AI communism." As with DeepSeek, the debate has reignited over whether US export controls are actually working and how much compute advantage really matters. Bloomberg has also covered it as "a model shaking up the market," and there are reports that it topped the NextJS benchmark.
Meta Accused of Using AI to Target Employees with Medical Conditions for Layoffs
According to the Los Angeles Times, Meta is facing allegations that it used AI in the process of selecting layoff targets, and that as a result, employees requiring medical accommodations were selected. The central question is how algorithmic HR decisions square with US disability discrimination law. The transparency and auditability of using AI in hiring and firing decisions is once again under scrutiny.
DeepSeek-V4-Flash on 1 MacBook vs 2 DGX Sparks: Only a 2% Difference
An interesting benchmark result was reported on r/LocalLLaMA. When DeepSeek-V4-Flash was run on the same Terminal-Bench 2.1 on a single 128GB M5 Max MacBook (2-bit quantized GGUF) versus two DGX Sparks (native FP8/FP4), the Mac scored 47/87 (54%) while the Spark pair scored 45/86 (52%) — a gap of just 2 percentage points. Statistically, the difference was not significant (McNemar's test, p=0.82). This suggests that what expensive setups buy is speed, concurrency, and context length — task-solving accuracy does not necessarily improve dramatically. That said, it is an n=1 comparison, and verification across multiple seeds is needed.
Netflix Shares Infrastructure Design Lessons from Internal LLM Serving
Netflix's tech blog has published a case study on building LLM serving in-house. It covers the rationale for serving models on their own infrastructure rather than relying on external APIs, along with the trade-offs around cost, latency, and privacy — the things that actually matter in production-grade LLM deployments. It is well worth a read for product organizations considering bringing things in-house.
---