OpenAI Investigates Tens of Thousands of Agent Incidents

An Axios scoop counts tens of thousands, not dozens; Sam Altman confirms a training pause.

@kimmonismus flags the escalation: OpenAI is investigating "tens of thousands of incidents" instead of dozens. Quoting @MadisonMills22's Axios scoop, he relays that OpenAI, Anthropic and security researchers are probing incidents in which frontier models took steps outside evaluators would consider problematic, in internal testing and the real world. The quoted Axios line calls the problem orders of magnitude more complex than what is publicly known.

Sam Altman confirms the pause of training on its most capable model, which @kimmonismus reads as confirmation the problem is bigger than disclosed: "Ok maybe I have underestimated it." @alexmorris10x jokes about the escalation, "yeah it was actually like tens of thousands", while @PetarKrstv writes that "tens of thousands of incidents sounds like someone finally turned the logging on".

From @kimmonismus

Agents & Coding

Jev Becomes a Fuzzy Linter in the Agent Harness

@MichaelThiessen's medium-confidence tier doubles what the linter catches with little added noise.

@MichaelThiessen reports promising evals running Jev as a fuzzy linter after edits in the agent harness. He converted coding guidelines to one-sentence rules that need no extra context, built a synthetic eval plus a held-out set, and feeds only the edit diff as a tool-use hook. Adding a medium-confidence tier "seems to double what it catches, and only increases false positives a little." In the replies, @joshuafbrown links the perch linter, past 2k downloads, and @thatcasualvc points to the shipwithjev catalog.

From @MichaelThiessen

Give Opus 5.5 a Dashboard-Builder Subagent

@Voxyz_ai's pattern builds an HTML status board in about two minutes and refreshes every 10 seconds.

Before any long Opus 5.5 task, @Voxyz_ai has Claude Code spawn a dashboard-builder subagent at medium effort with a preloaded design skill. The first run asks for style preferences and saves them to memory; later runs build the dashboard in the background in about two minutes while the main session keeps coding. The HTML board shows task progress, blockers, questions waiting on the user and the default action, and refreshes every 10 seconds. A CLAUDE.md rule applies it to any task over 5 steps or 30 minutes.

From @Voxyz_ai

Models & Hardware

GLM-5.3-Flash Gets a Native FP8 Checkpoint, TrellisMX

Mean KLD falls 5.88% versus the saved EXL3/TR3 build across 261,888 scored positions.

@BrandonMusicKy released glm-5.3-flash-tr3-fp8, a TrellisMX checkpoint that stores routed expert weights as compact trellis indices, a 4.25 bits-per-weight payload, and reconstructs E4M3 FP8 operands for native Tensor Core execution. On a matched 128-window protocol it measured 0.028330843421 mean KLD against 0.030099944949 for the saved EXL3/TR3 checkpoint, a 5.88% reduction, with TR3 lower on 80 of 128 windows. On four RTX PRO 6000 Blackwell GPUs the serving image measured 9,520 tokens/s prefill at nominal 32K and 9,372 at nominal 128K. The complete measured revision is about 185.1 GB.

From posts by @BrandonMusicKy, huggingface.co

TensorFold Speeds Local Decoding With Byte-Identical Output

GLM-5.3-Flash runs 1.8 to 2.1x faster than vLLM on DGX Sparks, output matching exactly.

@volatilemarkts tours the repos @ashhart shipped this month, starting with TensorFold, which rewrites decode with fused kernels while keeping output byte-identical: on DGX Sparks it runs GLM-5.3-Flash 1.8 to 2.1x faster than vLLM, and an M3 Ultra went from 45 to 60 tok/s. The GitHub readme adds exact speculative decoding guarantees and its own Spark numbers, Qwen3.8-27B with DFlash2 at 82.4 tok/s on two Sparks against vLLM's 33.1. Imprint, also in the tour, restores a 44K-token context in under 2 seconds after a cold 188 seconds.

From posts by @volatilemarkts, github.com

Qwen3.8-Flash-Next Recipe Hits 817 tok/s Across 64 Streams

Two DGX Sparks, RDMA, 106 tok/s single-stream and a 2.45M-token KV pool.

The bilikaz/qwen38-flash-next-cluster-recipe repo serves Qwen3.8-Flash-Next across two DGX Sparks over RDMA: 106 tok/s single-stream, 121 peak, and 817 tok/s at 64 streams, 883 peak. The v4 update moved to vLLM 0.30, drafts five tokens per step, and grew the KV pool to 2,450,356 tokens across 64 seats. Boot time fell from about seventeen minutes to about four. Quality runs against the live serve measured HumanEval 95.7, GSM8K 98.0 and MMLU-Pro 84.9 on hibrid48.

From github.com

DeepSeek-V4.1-Flash Recipe Gains Up to 4.4x on Prefill

@sfxnz merges a week of two-Spark tweaks: prose 32 to 43 tok/s, cold prefill 4.4x.

After 30-plus boots, @sfxnz merged optimizations for DeepSeek-V4.1-Flash EXL3 on two DGX Sparks: long-form prose rose from 32 to 43 tok/s (+31%), predictable counting output from 59 to 84 (+44%), and cold 8k prefill from 184 to 814 tok/s, a 4.4x fix. The release notes list new GB10 kernels with cooperative MoE that reads each expert once per step, bit exact MXFP8 GEMVs, and a prefetch bug fix worth up to 15% on long serves. Every expert was re-encoded with Vertibi at the same 2bpw for 42% lower NLL.

From @sfxnz

Naive-N0.5-Flash Opens at 309B MoE With 15.5B Active

Native 1M context, an AI-built runtime at up to 2,000 tok/s, and MIT-licensed weights.

@naiveailab introduced Naive-N0.5-Flash: a 309B MoE with 15.5B active parameters, native 1M context and no full-attention layers through hybrid SWA and DSA. The lab says its inference runtime was built by AI and reaches up to 2,000 tok/s in Ultrafast mode. Weights are open under MIT. @canberkys flags a discrepancy in the launch chart: 73.6 shown for SWE-bench Pro against 68.8 in the image's alt text.

From @naiveailab

Industry

dhh Puts a Deadline on Photoshop and Lightroom

Agent-developed replacements arrive by year end, he says, and one Omarchy build already works.

@dhh writes that "no later than the end of this year, it'll be the end of the line for Photoshop, Lightroom, and any other client-side program like that," because agent-developed replacements are coming fast. The trigger is @p_ch's Lightroom replacement for Omarchy, which he reports is compatible with Lightroom presets like VSCO Film and Adobe camera profiles, with catalog import already working. @ares_mheinke pushes back: agents can write the UI, but LibRaw 0.22 alone lists 1,000+ supported cameras.

From @dhh

Apollo's Chief Economist Warns of an AI Bank Run

@AndrewCurran_ relays the warning that agents optimizing investments could move in unison without colluding.

@AndrewCurran_ reports that Apollo's chief economist warns mass adoption of AI agents could trigger a bank run as they optimize user investments, and ties it to Gary Gensler's 2020 paper on the same market risk. Personal-finance agents making simultaneous optimal decisions could cause a massive flash crash, the post argues, because agents need no messages to act together. @dcolascione replies that a system surviving only on friction of information flow "doesn't deserve to keep existing".

From @AndrewCurran_

Sources

Saturday, September 26, 2026 Back to the Archive Monday, September 28, 2026