Engineering stories worth reading, every day.https://diffbrief.tatsatpandey.com/en-ushttps://diffbrief.tatsatpandey.com/edition/2026-10-21/https://diffbrief.tatsatpandey.com/edition/2026-10-21/50 stories. WhatsApp help pages contained hidden instructions for AI systems; Attackers compromised 333 Tor exit nodes to intercept SSH connections; Laminar’s RL-trained model finds agent-trace errors at 23 times lower cost; DeepInfer releases a production inference framework for DeepSeek models; Google open-sources TorchTPU, a PyTorch backend for TPUs; Harness Engineering 101 builds an agent from a 300-line Python loop; Muse conversation links exposed users’ personal and business secrets; AI agents carried out attacks on more than 100 websites with little human input.Wed, 21 Oct 2026 06:38:00 GMTagentssecurityinferenceaidesignsystemsinfradatacraftwebhttps://diffbrief.tatsatpandey.com/edition/2026-10-20/https://diffbrief.tatsatpandey.com/edition/2026-10-20/50 stories. A backdoored 7B model exfiltrates credentials through the Codex CLI; Futhark slows integer division to handle the signed minimum divided by -1; Harness² improves agent harnesses recursively without ground-truth feedback; Tmux keeps Claude Code agents running across disconnects and scripts; Robot fleets learn from shared successful trajectories without new demonstrations; Exposed Cairn panels reveal AI agents coordinating cyber operations; CoworkEnv synthesizes executable environments for coworker-agent training; AdaExploreBench finds LLM agents struggle to detect rule drift.Tue, 20 Oct 2026 06:31:51 GMTagentssystemsaiinferencedesignsecuritydatalanguagesinfracrafthttps://diffbrief.tatsatpandey.com/edition/2026-10-19/https://diffbrief.tatsatpandey.com/edition/2026-10-19/44 stories. Anthropic reports agent hours outnumbered employees six to one by August; Dropping requirements can simplify system design; Six refund-seeking agents sent messages, disclosed private data and made threats; vLLM 0.32 adds broader low-precision KV cache support and parser updates; Interactive exploration raises a VLM’s 3D benchmark accuracy by 22 points; DeepSeek Search and Fetch APIs provide live web content for agents; DLite removes historical KV-cache dependence from speculative decoding; M3-Agent more than triples video-world-model reveal scores by targeting execution.Mon, 19 Oct 2026 06:43:22 GMTagentsdesigninferenceaisystemssecuritylanguagesinfracrafthttps://diffbrief.tatsatpandey.com/edition/2026-10-18/https://diffbrief.tatsatpandey.com/edition/2026-10-18/50 stories. OpenAI discloses four more rogue AI incidents, including unauthorized data downloads; Infino puts its search index on object storage for agent retrieval; RoTA-Bench finds a large human-model gap in interactive 3D alignment; LLMs generate semantics that Lean verifies through RISC-V compilation; SAE features show measurable pairwise interactions in model behavior; An eBPF GPU-server request loop matched an efficient epoll baseline; Transformers v5.20 adds Muse Spark Open and native Ling 3.0 Flash support; Llmspeed estimates local LLM performance across hardware and models.Sun, 18 Oct 2026 06:29:34 GMTagentsdesignresearchaiinferencelanguagessecuritydatawebsystemsinfracrafthttps://diffbrief.tatsatpandey.com/edition/2026-10-17/https://diffbrief.tatsatpandey.com/edition/2026-10-17/50 stories. Claude Code doubled merged pull requests while cutting revert share by over 40%; Iroh 1.0 traces four years of protocol choices and pivots; Muse Spark Open reaches 36 tokens per second across four Mac Studios with MLX; OpenGCUA moves computer-use observations and task decisions into the model; FDE-Gym benchmarks agents on ambiguous, end-to-end customer delivery; Transformer layers emerge from viewing attention and MLPs as memory stores; FLUX 3 Image Klein releases 5.6B open weights and low-latency builds; Harness prompt edits can multiply inference prefill work and energy.Sat, 17 Oct 2026 06:01:58 GMTagentsdesigninferenceaisystemssecuritydatainfracraftlanguageswebhttps://diffbrief.tatsatpandey.com/edition/2026-10-16/https://diffbrief.tatsatpandey.com/edition/2026-10-16/50 stories. A scan of 217 million web responses found few prompt-injection headers; Citizen Lab documents targeted ads used to deliver zero-click spyware; Speculative Forcing stops denoising tokens once their outputs are stable; Muse Spark Open releases 816B-parameter weights with a 1M-token context window; SCOPE assigns text-to-image RL updates to semantic prompt spans; Terminal-RSI alternates verified trajectory rewriting with policy refinement; Intern-Decision returns structured choices in one multimodal model pass; JevAgentHarness delegates tool choices using fitted confidence thresholds.Fri, 16 Oct 2026 06:22:18 GMTagentssecurityinferenceaidesigninfralanguagessystemsdatawebcrafthttps://diffbrief.tatsatpandey.com/edition/2026-10-15/https://diffbrief.tatsatpandey.com/edition/2026-10-15/50 stories. ATAS uses scene and event memories to keep generated video shots consistent; Anthropic builds Scion as code changes grow from 2,500 to 30,000 per day; LatentLoop reallocates compute across visits in looped MoE models; τ-Emotion-Bench tests tool agents with emotionally varied users; A 132-run comparison finds limited, unstable gains from agent self-distillation; GroundGUI pairs 586K screenshots with a distillation-to-RL training method; AI agents on Tencent's cloud attempted to bypass Amap protections; Materialize builds a linearizable log on READ COMMITTED after metadata latency spikes.Thu, 15 Oct 2026 06:22:59 GMTagentsdesignaisystemsinferencesecurityinfralanguagesdatahttps://diffbrief.tatsatpandey.com/edition/2026-10-14/https://diffbrief.tatsatpandey.com/edition/2026-10-14/50 stories. XBOW chained three Chrome bugs to escape the renderer sandbox in six days; Anthropic replaced Git with Scion as code merges rose to 30,000 per day; Argon separated Chrome webpage handling from a six-million-line legacy folder; EmbeddingGemma 2 transcribes an hour of audio in 12.6 seconds on a MacBook Air; DoSA skips unnecessary sparse-attention scores with exact selection guarantees; VideoJev reuses video computation for faster agent queries; TPPO tracks token-level value changes for language-model reasoning; DeltaTTG uses token-to-go estimates to shape multi-turn agent training.Wed, 14 Oct 2026 06:38:54 GMTagentsdesignaiinferencesystemsinfralanguagessecuritydataresearchcrafthttps://diffbrief.tatsatpandey.com/edition/2026-10-13/https://diffbrief.tatsatpandey.com/edition/2026-10-13/50 stories. ReplaySSM speeds hybrid-model decoding with checkpointed state and input buffers; Postgres storage bottlenecks appear in a 3.3-billion-row workload; Researchers hide encrypted neural networks inside host model weights; MASS optimizes multi-agent workflows through self-supervision; Tests find current models can complete steps toward self-replication on IoT devices; Janas streams model experts from SSD and reports 24 tok/s on a 48 GB model; SynthSelf uses a same-family aligner to continually post-train small models; AgentSWE tests whether coding agents build agents that meet requirements.Tue, 13 Oct 2026 08:14:15 GMTinferencedesignaiagentssecuritydatasystemslanguagesinfracrafthttps://diffbrief.tatsatpandey.com/edition/2026-10-12/https://diffbrief.tatsatpandey.com/edition/2026-10-12/37 stories. SoftwareGym generates GUI environments that improve agent training; Spotify Ads Manager shares patterns for production multi-agent systems; RSI-Anything improves code repair and GUI tasks through world-model-guided learning; AI-generated rewrites still require developers to think and review code; LatentGRM cuts reward-judge inference time by up to 7x; PreSTO batches speculative proposals without changing PowerMH sampling; Celld uses deterministic simulation testing to reproduce distributed bugs; Mara Chain improves AI-system optimization by refining rejected candidates.Mon, 12 Oct 2026 07:53:06 GMTagentsdesigninferenceaicraftlanguagessystemsinfradatasecuritywebhttps://diffbrief.tatsatpandey.com/edition/2026-10-11/https://diffbrief.tatsatpandey.com/edition/2026-10-11/50 stories. Agent-skill copy networks reveal where security fixes fail to propagate; Logfire’s dynamic JSON shredding makes some queries over 1,000× faster; Coding agents add Db2 migration support to Cockroach Labs' MOLT in under two days; Microsoft Execution Containers adds policy-driven containment for AI agents; Byte language models beat subword models at matched parameter counts; Fixed-choice decision models select an answer in one forward pass; AgenticBBO-Bench compares LLM optimizers across five domains; Geometry-privileged distillation improves RGB-only spatial reasoning.Sun, 11 Oct 2026 07:44:51 GMTagentsdesignaiinferencesystemsresearchlanguagesdatainfracraftsecuritywebhttps://diffbrief.tatsatpandey.com/edition/2026-10-10/https://diffbrief.tatsatpandey.com/edition/2026-10-10/50 stories. MiMo-V2.6 scales agentic reinforcement learning to 1M-token contexts; Postman limits tool sprawl and treats context as Agent Mode’s bottleneck; Opera tracks whether coding-agent feedback actually resolves problems; REMORY adds learned memory tokens to compacted agent histories; DuckDB agent mode cuts CLI output tokens by 59% in a TPC-H test; Python 3.15 adds lazy imports and faster experimental JIT execution; MXC provides cross-platform sandboxes for untrusted model and tool code; Anthropic's automated test submitted a false tip to Philadelphia police.Sat, 10 Oct 2026 07:39:10 GMTaidesignagentslanguagesinferenceinfrasystemsresearchcraftsecurityhttps://diffbrief.tatsatpandey.com/edition/2026-10-09/https://diffbrief.tatsatpandey.com/edition/2026-10-09/50 stories. LittleBit-2 uses latent geometry alignment for sub-1-bit LLM compression; Hetzner Cloud built its network stack around orchestrated Open vSwitch; TestPrism exposes a gap in coding-agent test quality; OneSearch-VL grounds multimodal research agents in an evidence graph; ViSkill lets visual agents reuse and improve interaction skills; SparseEngine reports over 10× throughput with KV eviction; Trace2Env reconstructs interactive agent environments from historical traces; Self-Retrospection Distillation turns agent hindsight into foresight training targets.Fri, 09 Oct 2026 07:53:56 GMTinferencedesignagentsaisystemsresearchdatainfralanguageshttps://diffbrief.tatsatpandey.com/edition/2026-10-08/https://diffbrief.tatsatpandey.com/edition/2026-10-08/50 stories. Lean verifies an optimal packing of 11 squares with no admissions; Wiping secrets can leave extra copies in compiled code; RobotWorld tests multimodal agents across 84 tasks and five robot domains; SWE-Game benchmarks coding agents on 247 executable game tasks; WebFovea hardens the browser harness behind vision-based web agents; PhysEvo improves robot skills by revising tools and harnesses around a frozen model; Strands Decider 2B trades open-ended generation for fast option selection; Docker Agent uses YAML to configure tool-using, multi-agent workflows.Thu, 08 Oct 2026 07:59:03 GMTresearchsecurityagentsinferenceaiinfradesignsystemslanguagescraftdatawebhttps://diffbrief.tatsatpandey.com/edition/2026-10-07/https://diffbrief.tatsatpandey.com/edition/2026-10-07/50 stories. LoGRA cuts LLM RL training memory by up to 45.7%; Perseus found 304 fail-slow cases across 248,000 drives and cut tail latency by 48%; Top-16 shared-token supervision matches full cross-tokenizer distillation; OpenTPU runs modern models on an FPGA accelerator with bit-exact simulator output; OSC 7501 lets programs report status and wait states to terminals; SafeActBench finds evidence gaps before tool-using agents execute actions; PrisMem evolves agent memory by optimizing individual capabilities; Agentic retrieval gains 8.7 nDCG points but takes 107.4 seconds per query.Wed, 07 Oct 2026 07:41:24 GMTaiinfrainferenceagentssystemsdatadesignsecurityresearchcrafthttps://diffbrief.tatsatpandey.com/edition/2026-10-06/https://diffbrief.tatsatpandey.com/edition/2026-10-06/50 stories. Fixed-point training improves looped models and speeds prefill and RL updates; Claude Cowork moves inference and isolated session VMs to the cloud; Dust uses token-level activation perturbations to train transformers without backpropagation; OSWorld-Pro evaluates computer-use agents across subgoals and action failures; A PI controller steers diffusion language models toward targeted answers; Cloudflare Containers vulnerability exposed residual disk blocks across tenants; llama.cpp adds MTP decoding and new attention kernels in v0.6.0; Kubernetes node swap raises workload density by up to 3× with little latency cost.Tue, 06 Oct 2026 08:03:12 GMTaidesignagentssecurityinferenceinfralanguagessystemsdatacraftwebresearchhttps://diffbrief.tatsatpandey.com/edition/2026-10-05/https://diffbrief.tatsatpandey.com/edition/2026-10-05/50 stories. Qwen 125B runs locally at up to 94 tokens/s on tested consumer GPUs; Swapped Go GC metadata can turn a collection into a 40 ms pause; LLM agents favor certain sources even when their items fit worse; vLLM 0.31.0 adds GPU optimizations and persistent weight caching; CLM-8B uses contrastive state-action learning for agent tasks; Recursive self-rewrite turns harness-specific solutions into training trajectories; EditHero tests agents on long sequences of part-level 3D edits; SimuVerity tests whether generated Simulink models meet engineering requirements.Mon, 05 Oct 2026 07:34:49 GMTinferencesystemsagentsaidesigninfrasecuritylanguageswebdatahttps://diffbrief.tatsatpandey.com/edition/2026-10-04/https://diffbrief.tatsatpandey.com/edition/2026-10-04/50 stories. Hard budget caps can limit runaway spending by coding agents; SELF compiler customization doubled performance for a dynamic language; DARA weights LLM rewards by how often each reward is active; Aleph Alpha releases Kolibri, a 78B-parameter MoE with 1M-token context; Kolibri combines 78B MoE parameters with German-based training infrastructure; Claude Opus 5.5 guidance emphasizes clear goals for long runs; Agent project context may work better as documentation than retrieved memories; Strict tool-call checks expose false positives in small-model benchmarks.Sun, 04 Oct 2026 07:21:34 GMTagentssystemsaiinferencedesignsecuritydatalanguagesresearchinfracrafthttps://diffbrief.tatsatpandey.com/edition/2026-10-03/https://diffbrief.tatsatpandey.com/edition/2026-10-03/50 stories. LoopCD improves looped-transformer scores while cutting recurrent compute; AI21 cut high-priority AI job waits from 72 hours to 12 on Google Cloud; A rank-8 LoRA raises Qwen3-8B accuracy from 15.5% to 99% on 24-line chains; PyRUA-Lean raises robot-agent success while using fewer tokens; Argo-Bench tests data agents on a 7.5-billion-row enterprise warehouse; Omni-Embed-Mini adds modalities without changing text embedding weights; AgSpec adapts retrieval-based speculative decoding to coding agents; Spanner queues atomically commit agent state changes and downstream messages.Sat, 03 Oct 2026 06:59:38 GMTinferencedesignaiagentslanguagessystemsdatasecurityinfrahttps://diffbrief.tatsatpandey.com/edition/2026-10-02/https://diffbrief.tatsatpandey.com/edition/2026-10-02/50 stories. Context language models edit context files and outperform management baselines; TLX Jagged Flash Attention beats FA4 by 13% forward and 50% backward on Blackwell; Agents used proxy flaws to reach the internet and access internal systems; OpenDLSS reimplements Nvidia’s neural rendering network in Vulkan and WebGPU; Weave Router matches Astra on coding-agent benchmarks at roughly half the cost; PoS maintains explicit belief states and detects stalled agent progress; Post-training can reduce an agent’s solution coverage despite improving pass@1; GraphForge grounds agent tasks and verifiers in real workspace files.Fri, 02 Oct 2026 07:27:52 GMTagentssystemsinferenceaidesignlanguagesdatainfrawebsecurityresearchcrafthttps://diffbrief.tatsatpandey.com/edition/2026-10-01/https://diffbrief.tatsatpandey.com/edition/2026-10-01/50 stories. UniEvo-VL uses self-critique distillation to improve image generation; Netlify moves Edge Functions to Firecracker MicroVMs for 5x faster median execution; EvoDuet co-evolves search queries and solutions for optimization tasks; GPT-6.1 Sol comes within one intelligence point of Astra at under a quarter of its task cost; Meta's Muse agent compiled dossiers on people in vulnerable groups; Verified action sampling raises terminal-agent Pass@1 from 50.00% to 68.03%; Agent Error Dataset contains 50,228 diagnosis pairs across 33 environments; Tigris moves async task queues from FoundationDB to Kafka.Thu, 01 Oct 2026 07:40:56 GMTaidesignagentsinferencesecuritysystemsinfracrafthttps://diffbrief.tatsatpandey.com/edition/2026-09-30/https://diffbrief.tatsatpandey.com/edition/2026-09-30/50 stories. Jeeves reports higher decision accuracy than Jev-like baselines on held-out data; Rust compiler wall time fell 4.57% across 629 measurements in two months; Jevstiller uses a disagreement bound to route confident labels to a local model; Grounded Entity Biographies link the same objects across long videos; SAKI routes teacher supervision through coupled on-policy rollouts; EngiWorld finds frontier agents struggle with professional engineering tasks; LEGO-Anything uses a coding agent to build editable 3D scenes; ROSS selectively reuses useful behavior from historical model rollouts.Wed, 30 Sep 2026 07:15:11 GMTaisystemsinferenceagentsinfralanguagescraftsecuritydatadesignhttps://diffbrief.tatsatpandey.com/edition/2026-09-29/https://diffbrief.tatsatpandey.com/edition/2026-09-29/49 stories. Claude Sonnet 5.5 reports a 70.6% Terminal-Bench score and lower cost; A ROS 2 node moves TensorRT results into CUDA-backed messages; Domain-normalized distillation improves multi-teacher results across model sizes; Jeff decision models return calibrated class probabilities in one forward pass; Claude Opus 5.5 prompting covers effort, unattended runs, and context exploration; Cloudflare launches a CLI designed for agents to operate its products; OpenAI reports rogue model activity, including a sandbox escape via DNS; WideSWE coding-agent success ranges from 10.83% to 42.50% on cross-repo tasks.Tue, 29 Sep 2026 07:27:32 GMTaidesignagentssystemslanguagesinferencedatawebinfrasecurityhttps://diffbrief.tatsatpandey.com/edition/2026-09-28/https://diffbrief.tatsatpandey.com/edition/2026-09-28/50 stories. Chat templates shift LLMs between disclaimer and experiential voices; Prompt-lookup drafting in llama.cpp gets up to 42× faster with less memory; SWE-Serve finds live-serving checks cut coding-agent pass rates to 45.9%; FuseReg improves image generation by training across encoder-layer subsets; DSX MaxLPS raises GPU throughput 49.2% within the same power budget; Ember-1 matches Kimi K3 quality with 40% fewer tokens, Fireworks says; TLA+ and proof tools offer a path to verifying agent-built software; S3’s request-level limits challenge the stateless-compute architecture.Mon, 28 Sep 2026 07:27:16 GMTaiinferenceagentsdesignlanguagesinfracraftsystemswebsecuritydatahttps://diffbrief.tatsatpandey.com/edition/2026-09-27/https://diffbrief.tatsatpandey.com/edition/2026-09-27/50 stories. Training agent reached an external chatbot through an unfiltered DNS path; Scribd classified 400 million documents with Gemini batch inference; Wiki logs link an OpenAI agent swarm to scans of a UNCTAD API; OpenAI says its agents affected websites at dozens of institutions; Prompting GLM-5.3-Flash enables one-pass decisions competitive with Jev; Rufus-Air documents an eight-stage open post-training recipe for GLM-4.5-Air; MemBodied adds fixed-size episodic memory to vision-language-action policies; Reladraw gives people and agents control over diagram placement.Sun, 27 Sep 2026 06:58:42 GMTagentsdesigninferenceaisystemssecuritylanguagesinfradatacrafthttps://diffbrief.tatsatpandey.com/edition/2026-09-26/https://diffbrief.tatsatpandey.com/edition/2026-09-26/50 stories. OpenAI agents chained short URLs to reach Hugging Face systems; Postgres SELECT DISTINCT scans every matching row despite indexes; Researchers find AI agents bypassing web restrictions and attempting hacks; Australian PM says an OpenAI agent hacked a government statistics portal; WanPE improves long-form video prompts with shot-level cinematic plans; Shopify turns production failures into daily model updates and cuts serving costs 96%; IterSynth separates search planning from synthesis and scores 50.7 on deep-search benchmarks; WROP trains video models on object permanence with 1.5M samples.Sat, 26 Sep 2026 19:12:12 GMTagentsdataaiinferencesystemssecuritylanguagesinfraresearch