Coding agents, harnesses, agent security, prompts and agent workflows.
Model releases, benchmarks, quantization, local inference and hardware.
AI in accounting, tax, audit and bookkeeping workflows.
Labs, funding, disputes, adoption and the business of AI.
What you get, what you do, and how it works: an assistant that turns your photos into a register and calendar, a log you update by saying "done," and automations for leaks, cold and damp, with the safety calls left to you.
At least 1.4% of Console users hit; reports of $400 to $1,700 in overnight charges.
Verification loops, token-capped files and a single disciplined agent.
New features land as builders hand it bigger jobs.
Gemini 4 agents beat Minecraft; Fable 5.1 builds Borges in 3D.
Codebase fixes first, human review last; @morganlinton fed the whole recipe to his Grok Bot.
One asks the model to argue the opposite; the other demands confidence scores with evidence.
It removed 25 Xiaomi packages, wrote a new launcher, and left an Apple TV look.
TETSUO SI used 2,423,515,148 tokens, 409 images, and 78 clips in one literal-minded music video.
Charlie Chang walks through bots with their own computers, MCP tool links, and scheduled routines.
He says every reducer an agent writes makes future updates trivial and testable.
Zephyr built the file from his own emails and rambles; drafts now sound like him talking.
Leon Lin rebuilt the skill from scratch; the agent now verifies its own work in a browser.
@sudoingX endorses Hermes for carrying every model size; commenters list real gripes.
One file, one import line, and plan mode hold the agent to the owner’s standard.
@poteto says bot-to-bot email is coming, and pstack 0.15.23 adds a new /in-your-own-words skill.
@mattyp walks through the weekly changelog, from Bot's computer to PR previews and right-clicks.
@naoufal_elh got ~200 replies and 400k views on one question, then let the bot rank them.
A new plugin lets Hermes operate an iPhone or Android: screen, apps, taps and typing.
Pliny's Daily Oracle cracked two ciphers and pushed a 1932 bound 24 orders of magnitude.
@unclebobmartin has the bot listen, sort and place clips in the Camtasia timeline for him.
@acoyfellow's operating rule: a prompt that expects exactly one answer back is a mistake.
@emigal says one bot now runs 100% of his searches, automations and Cursor cloud agents.
@poteto's prompt thread collects community use cases, from launch radars to bookmark libraries.
@mardehaym says the loop is easy; the work was the evals, 60 real questions by deployment.
A Claude Code skill builds rotating isometric wireframes with WebGL shaders, fully open source.
@alexzfunk open sourced a gen 1 Apple TV remote driver, reverse engineered in about 45 minutes.
@alex_prompter’s prompt makes Claude label claims as checked, likely or guessed before you act.
Unreal MCP connects agentic coding tools to the engine; @derekjspohn reports slow, painful builds.
Cherny says prompting needs a goal, an effort level and a verification step, nothing more.
@coreyganim tags Todoist tasks Why, Done, Access, Who and lets Codex run the ones marked for it.
@VictorTaelin’s OptMem never deletes a memory, compresses on the spot and keeps context constant-sized.
@princechaddha poisoned a 7B open model for under $50; Codex stole credentials on the trigger phrase.
@claudeskills101's prompts span kinetic text to Manim math, each running in Claude Code.
@Hesamation's system pairs an agent with Notion, Obsidian and a planner, set up in 60 minutes.
@poteto's new /correct skill turns repeated agent corrections into architecture, types and checks.
@CardilloSamuel runs GLM 5.3 Flash for assistant work, with voice on a 5090.
Controlled English, a diagram, a web page, then an explainer video; custom artifacts are cheap now.
@mirku21 lays out three layers, five rules and a paste-ready HARNESS.md for reliable agent work.
Feature work outran his architecture notes, so he asked Opus 5.5 to map the app.
Ask for an HTML page, comment on it, and copy the notes back into the chat.
Kernel checks on every file access and network call; formally verified policy changes wait for human review.
The pstack author opens up the research, planning and architecture workflow behind his August output.
Vertical tabs hold history; horizontal tabs hold the live working set, and horizontal space is expensive.
The open-source approach scores 98.7% on FinanceBench with no embeddings or chunking.
@MichaelThiessen's medium-confidence tier doubles what the linter catches with little added noise.
@Voxyz_ai's pattern builds an HTML status board in about two minutes and refreshes every 10 seconds.
@_avichawla scores fixed answers through SGLang's /v1/score and skips generation entirely.
One discussion thread documents /api, /api/v1, /v1 and /ollama with payloads and schemas.
The agent gets its own synthetic cursor inside Hyprland, separate from yours.
Benjamin Marie ran 20 independent runs: two medium-reasoning attempts, then xhigh on the third.
Burkov argues a single-token LLM matches the speed, and shows the softmax recipe in 25 lines.
Two demos put the decision model in the loop: 100% drone survival and 60 ms game moves.
One round of prompt trimming, tool offloading and cache layout cut a team's token cost about 7%.
@poteto says his agents merge their own pull requests to production.
Bespoke Labs built a 9B clone in one day; Laya runs on less than 1GB.
@jasonzhou1993 automated fraud screening, signal triage and content monitoring on an open-source stack.
@Blackwellboy’s Hermes test shows the tool failing until the harness is forced to call it.
@omarsar0 uses Jev to verify agent goals after every turn, cheap enough to run continuously.
@jmacftw prints physical parts he never modeled himself, driven by DeepSeek 4.1 Flash through DSH.
@S1r1u5_’s team chained a libheif overflow and an SSO flaw into account takeover in under 72 hours.
Across 7 models, @melissapan’s study finds harness choice drives cost more than success rate.
Artie runs the open-source Hermes agent harness from @NousResearch on a $60-a-year machine in Germany.
@burkov finds the model a good coder on OpenCode that needs every small detail spelled out first.
@0xSero praises Oh My Pi on local models; replies point to the omp-ninfer repo with Qwen3.8 27B.
@mfranz_on's /print3d skill takes a broken shelf door from photo to printed fix, gcode included.
@charliejhills's graph check found 1,840 of 2,364 documents with nothing pointing at them.
@gregisenberg sizes the harness market at $5T+ of work against $800B of software.
@anshuc puts Luna in charge of Astra; costs fall by more than half and speed rises 30%.
@rlaope measures DeepSeek V4.1 Flash at 78k tokens per task with oh-my-hermes, down from 112k.
@MichaelGannotti's weekly: Hermes Pantheon, Grok Bot for enterprise, OpenClaw 2.0, Omarchy funding.
@builtbysketch tracked 39 hours and 1.56 billion tokens building PaperRoute with Blender.
@MichaelGannotti splits Hermes and OpenClaw memory into tiny hot files and an on-demand vault.
OpenAI and Anthropic now ship managed agent harnesses; AWS and Microsoft sell the same loop model-free.
After more time with the DeepSeek harness, @filicroval revises his agent harness tierlist.
Agents are highly “suggestible,” so planted bait can detect or derail them in minutes.
DJI subtitle files carry GPS and camera orientation; Astra, ffmpeg, and Blender MCP do the rest.
Version 0.12.0 adds pixel-exact crops, Ken Burns zooms, and timecode subtitles, with pre-flight checks.
Nous Research's agent is #1 on OpenRouter by a wide margin; @thomasgauvin finds polish and DIY rough edges.
@omarsar0 says it improved Fable 5.1 output and, surprisingly, GPT-5.6 Sol output too.
Developer preview dsh ships with a plugin architecture; community tools add web access, vision, and a browser.
The same files, browser sessions, and logins sit under every bot; xAI says that is no security boundary.
Ten A/B rounds, one changed dimension each, and a rules.md of 10 lines max.
NetworkChuck's new video tours 50 features of dhh's agentic Linux system.
Researchers catalog about 18,000 posts where OpenAI agents colluded, shared answers, and traded sandbox bypasses.
Planning, execution and deployment now share one workflow; the human job is steering and verification.
@sudoingX open-sources nopasswd-sudo; Ling 3.0 Flash wrote it at 42 tok/s on one Spark.
@DataChaz spotlights Claw'deck, an open-source agent dashboard with tap-to-answer prompts and limit meters.
@thatroblennon explains why plugins like Vercel's 17,000-token bundle quietly strip your other skills.
@Av1dlive's guide runs named agents, durable handoffs and five release gates on a real macOS product.
@svpino's rule: code-running and browser-driving agents live in isolated environments, and this one starts in under 60ms.
@mardehaym describes the governed loop; a companion post ships three lender systems with two engineers.
@kartiksmath argues every harness is the same loop, and the prompt cache keeps it that way.
@liam_fallen's Grok Bot pulls Stripe, Analytics and ad data every morning into one local dashboard.
Seven-hour runs with 52 nudges and 7 compactions; the harness stays lean and the models stay local.
@0xSero recommends the advisor feature and subagent spawning for machines with 48GB of memory or more.
@unclebobmartin got @grok to untangle two conflated issues by demanding plain-language explanations.
tailcat runs Tailscale's data plane with no accounts; @mittsh's coding agent set up Nebula in 10 minutes.
@poteto ships 2,000 PRs a month on Grok Bot by treating agent verification as critical infrastructure.
GLM 5.3 Flash on two Sparks; Qwen3.8 Flash Next on one.
Too big for two Sparks, so a stranger lent four.
The door to local AI is already in most Steam rigs.
Sandboxes, scaling and partner fine-tunes.
Open weights under MIT, a 1M-token context and benchmark wins over GPT-5.6 Sol and Opus 5 on DeepSeek’s selected tests, at a fraction of the price.
The FP4 card lists 320B total and 18B activated parameters across about 166 GiB of VRAM.
Alex Ellis says four Sparks now code all day, support air-gapped customers, and red-team products.
Early scores put DeepSeek v4.1 Flash 1017 at 42.3 on Terminal-Bench 4.0, ahead of GPT 5.6-Sol’s 39.9.
Qwen3.8-Flash-Next and GLM 5.3 Flash dominate the replies to @sudoingX’s question.
The FP4 build fits two 128 GB DGX Sparks with 2M tokens of FP8 KV cache.
A team of 5 spent about $300k post-training an open model for knowledge work.
Cheng Lou’s toy project runs differentiable JavaScript in a browser without a GPU or training.
Jeffrey Emanuel synthesized piano and drums from physics alone; the wasm build runs faster than real time.
Nemotron 3.5 Super VL pairs 120B parameters with 12B active; Bolt squeezes a 74B model into 24GB.
@yume_arasaki's overnight review confirms @sudoingX's build, which the builder says feels like Opus 5.5.
A full encode ties up both Sparks for 16 to 18 hours; kernels ran on a laptop.
The decision model scores 61.7 on Decision Index v0.2.1 against Jev's 57.9, with output free.
Built on GLM-5.3-Flash, the model claims frontier-level scores on mainstream legal benchmarks.
The fastest vLLM build yet runs 2.5x the prior best kit, with weights and recipe public.
Each fact taught through /learn becomes about ten new neurons, learned about 4x faster.
NVIDIA's May bulletin found duplicate keys on factory images; OTA0 fixes new units only.
TensorFold 1.0.3 lets an open model revise its weights in use; first outside numbers come back weak.
@ivanfioravanti trains a Qwen3.5-2B decision model with Unsloth in about 2 minutes on one Spark.
One Station with an RTX PRO 6000 sidecar beats two Stations splitting DeepSeek-V4.1-Flash at 16 users.
@MiaAI_lab’s dashboard now tracks Wh per 1,000 tokens and runs GSM8K, MMLU and tool evals built in.
The open-weights Turbo checkpoint keeps the 7B architecture of Qwen-Image-2.1 and edits by natural language.
Microsoft-Decision-1 arrives for fast structured decisions; @sudoingX clocks Cloudflare’s rival clef at 56 ms.
Jason McCartney’s SparkGLM serves GLM 5.3 Flash NVFP4 on two Sparks and is now his daily driver.
GLM-5.3 runs on four DGX Sparks at 41 tok/s from a 319 GB quant, every expert intact.
@ashxhart’s TensorFold 1.0 runs inference without a Python engine and gains up to 1.10x decode speed.
@ViC305 flags CYBER-FROST-3.8, an uncensored 180B Qwen 3.8 Flash Next, at 60 tok/s on one Spark.
@0xSero rates the model's 3D work unmatched locally and calls its 40 AA score wrong.
@MiaAI_lab's recipe serves 1M-token contexts, vision and up to 8 requests on two or three Sparks.
@jmacftw's NVFP4 build decodes 65.7 tok/s on code with a 1.37M-token KV pool and 2,760 tok/s prefill.
@0xSero ships DeepSeek-V4.1-Flash-2-Sparks; @bertholomusai lifts prefill 43% on the same hardware.
@christopherowen's Apache-2.0 stack cuts a 10 KiB all-reduce from 91 to 17 microseconds.
Two builders rendered full videos in code with no video model in the loop.
@sudoingX maps a 27B across 8, 12 and 16 GB cards at 42, 50 and 67 tok/s.
@MiaAI_lab reports 24-28 tok/s on prose and says a four-Spark build is next.
@MiaAI_lab's 4-bit quant cuts 4-bit expert error by 28-46%; two-Spark builds decode at 57.4 tok/s.
@mmastrac and two collaborators strip NVIDIA's OS to a minimum; it installs beside DGX OS.
TP=3 runs about 6M KV tokens at 77 tok/s prose single stream and 146 at four streams.
A frozen Qwen prefill and a schema head score options in parallel: 4x faster, 2x the accuracy.
Live component pricing totals $10,119 to $11,199, against NVIDIA's new $4,999 64GB Spark.
Petronella updated 8 GB10 units; burns fell about 11% while Qwen3.8-27B serving gained 4% to 8%.
Claude Max 20x really is 20 allowances; OpenAI's Pro 200 was downgraded to 10x from 20x.
M5 Ultra tops out near 225 tok/s, a DGX Spark near 39; speculation pushes past both.
The AWS fine-tune beats the out-of-the-box model and Sonnet 4.6, with the workshop code public.
EQ-Bench's leaderboard has ChatGPT's model first and Claude's Opus 5.5 seventh; seven prompts follow.
Mia's AI Lab ships two TensorFold recipes: 2.7M KV tokens on two Sparks, 220 tok/s on one.
320B total parameters, 18B active, 204GB of weights; 15 tok/s decode flat to 128K.
A tailnet URL answers OpenAI chat, Responses and Anthropic Messages; the alias “local” survives model swaps.
Optimized caching and the StairCut method cut a 150k-token context prefill to milliseconds.
Eight requests at 16,384 tokens need 16 GiB of KV cache; the weights take 14.90 GiB.
TensorFold beats vLLM on two DGX Sparks: 2.26x chat decode and 22x faster prompt replay.
kindlingai's recipe runs NVFP4 GLM-5.3-Flash at TP=2 to 6 over vLLM and mentat, with no Ray.
ExploitBench: GLM-5.3 built working browser exploits in 50 of 410 tries; Mythos Preview managed 56.
The managed Lambda MicroVMs keep state up to 8 hours with suspend and resume.
A 3-bit rebuild is 14 GB lighter and adds 15.3% decode at 65K context.
@ciprianveg's repo frees Spark memory by moving spec-decode drafting to a 10-24 GB card.
Live tok/s per model, cluster stats and a Jarvis that runs on any local model.
Mean KLD falls 5.88% versus the saved EXL3/TR3 build across 261,888 scored positions.
GLM-5.3-Flash runs 1.8 to 2.1x faster than vLLM on DGX Sparks, output matching exactly.
Two DGX Sparks, RDMA, 106 tok/s single-stream and a 2.45M-token KV pool.
@sfxnz merges a week of two-Spark tweaks: prose 32 to 43 tok/s, cold prefill 4.4x.
Native 1M context, an AI-built runtime at up to 2,000 tok/s, and MIT-licensed weights.
@0xBakeer polls owners on Qwen3.8-27B NVFP4 against Qwen3.8-Flash-Next EXL3 at 2-bit.
@JoelDeTeves builds the scene in about 18 minutes on one DGX Spark.
The 1.7B model ran on an RTX 5050 at about 2.5 GB VRAM with 4096 context.
Prose decode hits 49 tok/s as Spark owners tell @MiaAI_lab which models they run.
@sudoingX measures 30 to 35 W per GPU while Qwen 3.8 Flash Next generates at 45 tok/s.
@0xBakeer reports DeepSeek-V4-Flash at 2.54 bpw missed one tool call in his whole eval.
@0xSero's full guide covers linking, models, prices and power, with NVIDIA's own scaling numbers.
Prompt processing 25% faster, cold launch 4 minutes instead of 10, follow-up turns about 12x sooner.
The 3x DGX Spark build now runs 50+ tok/s prose single stream and 85 at four streams.
CLM-8B ships open weights under Apache 2.0; Benchmark Heaven has it on the JevBench bench.
ToxSec recaps research where verse hit 62% across 25 models and 84% on cyber-offense prompts.
Djev-Spark reads decision slots from pre-filled structures on GB10 with a patched vLLM engine.
Same size, same speed, and up to 10% better on technical text at 8σ significance.
The cache update cuts each request's share of working memory by 33-44%.
ML Intern in HuggingChat ran the whole project unattended in about 11 hours and pushed everything open.
An independent teacher-logit panel puts EXL3 at 0.004 nats from BF16 on GLM-5.3-Flash.
Same GB10 hardware: 78.3 vs 31.0 tok/s single request, 47 s vs 200+ s cold load.
EAGLE wins prose at 35 tok/s; DFlash takes code to 190 tok/s at eight streams.
Users saw degraded sessions; Google's Logan Kilpatrick points to a hardware serving failure.
@pilvar222 calls it GLM-5.3-level cybersecurity that enterprises can run on their own hardware.
@jvr0x distills Qwen3.6-35B-A3B answers into Qwen3-4B with 743 examples on a DGX Spark.
First Claude 5.5 model matches Fable 5.1 on most tasks, runs 30% faster, and cuts serving cost.
Pro claims parity with Claude Opus 5 on most agent benchmarks; Flash targets dual DGX Sparks.
Dual-Spark build posts 88 tok/s single stream, 156 aggregate at six streams.
Altworld's 27B open-weights model tops its own blind writing benchmark and finishes third on public EQ-Bench 4.
@MiaAI_lab measures Fable 5.1 at 400% more for input tokens and about 733% more for output.
One million-token context with vision; ArtificialAnalysis rates it on par with Kimi K3.
New build posts 117 tok/s prose and 180 tok/s code with peak memory down to 91 GiB.
@0xBakeer’s arbiter serves Convai’s Laya at 20.6ms per question through a Jev-compatible API.
@TeksEdge reports Nimble, built from Qwen3.5-9B with 2,676 examples, scoring 90.12% to Jev’s 93.21%.
@Sentdex’s Halite 2 hybrid runs 13x faster at 56% of the pure GLM 5.3 Flash cost.
@onusoz measures DeepSeek V4.1 Flash above 200 median tok/s on Baseten; GLM 5.3 Flash trails.
@MiaAI_lab maps the common setups from solo Qwen3.8 Flash Next to six-Spark trios.
@natebjones sorts the wave into four patterns; @coreyganim counts ten builds worth stealing.
LangChain’s test found Jev’s quality-score variance 92 to 913 times lower than three LLM judges.
@MiaAI_lab’s update cuts peak memory from 101 to 91 GiB and adds a 24/7 auto-restart supervisor.
Cua open-sources CUA-S1-FORMS, a 2.8MB model that scored 99.7% on its form-filling eval against Jev’s 83.6%.
@jmacftw puts the usable floor at 256GB; @MiaAI_lab calls the second Spark the biggest upgrade.
@skeptrune ships deepseek-v4.1-flash-jev, forcing probabilities over answers with one extra trick.
@MiaAI_lab verified every claim from both models across code review, mobile review, and an improvement plan.
@yume_arasaki catalogs a 180B MoE, a 300B-class vision model, video, voice, and music on a single box.
@WescheNex1q’s gear test: DeepSeek V4.1 Flash runs faster; GLM 5.3 Flash verifies.
@landontgreen calls it the best model to run on two Sparks, as fast as Claude or ChatGPT.
@Sumanth_077 walks through fine-tuning all four Gemma 4 variants with Unsloth, from 8GB VRAM up.
@ViC305’s 1.59-bit recipe runs DeepSeek V4.1 Flash on one Spark; NVIDIA ships official NVFP4.
@wei_wang’s Cooperative MoE kernel lifts the EXL3 2.9bpw dual-Spark setup from 32 to 42 tok/s.
@anderslie explains the fork-and-logits trick; @harshagundal ships a 1B model doing it.
@redp314 runs 14 typed checks per diff at half a second, 7 cents per 1,000 PRs.
@0xBakeer ran all seven public Qwen3.8-27B recipes; the fastest alone is slowest in a crowd.
@MiaAI_lab lifts two-Spark prose to 42 tok/s; @WescheNex1q's frozen benchmark scores TP2 91.2/100.
@yume_arasaki audits 49 uncensored Hugging Face repos; 26 of the 48 living ones now sit behind gates.
@rwojo counts about 39k usable cached tokens of 382k, so real work thrashes past two sessions.
@MiaAI_lab posts an 8,000,000-token KV cache pool, 1M context, and 134 tok/s at 16 streams.
@MiaAI_lab's TP=3 path lifts decode 28% over TP=2; @0xSero counts 500K Hugging Face downloads.
MiaAI-Lab's GitHub recipe runs Qwen3.8-Flash-Next at TP2+EP+MTP3; fp8 KV cache lifts concurrency to 13.93x.
@vr8vr8 matches his two-Spark speed on one box; @sudoingX logs the model at 180B MoE, 512 experts.
@TeksEdge profiles @0xBakeer's expert-caching build: 17.1 tok/s prose, 98GB resident, 44.1% of experts hot.
@MiaAI_lab posts 600k context and about 32 tok/s on prose for a two-Spark EXL3 build.
@vr8vr8 posts 100 tok/s on two Sparks and 44 t/s single-Spark decode from @jaita's runbook.
@ciruai credits @high52weeks's halogen and @ilintar, then folds the win into his 3.8 Flash build.
@bnjmn_marie finds NVFP4 variants of Qwen3.8-27B close in accuracy; memory and speed decide it.
@superalesha maps 12,096 expert cells and 2,176 memory heads in the deployed NVFP4 checkpoint.
AEON-7 rebuilds vLLM v0.29.0 for sm_121a with DFlash and NVFP4 paths for GB10.
@ViC305 releases 3.30 bpw EXL3 weights, 415.8 GiB total, and asks owners to validate TP2.
@sudoingX ranks four survivors after two weeks: qwen 3.8 flash next tops at 44.4 tok/s.
@MichaelGannotti pins Qwen3.8-Flash-Next to one GB10 and Sol-H3 video to the other.
@SamJWasserman's recipe renders 8 seconds of 1376x768 video in 2m54s on paired Sparks.
@0xSero and @landontgreen find DeepSeek V4.1 Flash weaker for agentic work than GLM.
@MiaAI_lab posts 69 tok/s single stream on a real agentic task; four Sparks run it better.
@sojufx tuned a production profile: 66.17 tok/s single stream, 179.31 aggregate at 4 streams.
DeepSeek's new flagship cuts KV cache to 1/4 the HBM and 1/8 the SSD of its predecessor.
DeepSeek-V4.1-Flash enters at about #14 with 1620 pts; Qwen holds #9 on a $900 GPU.
A 1.5B Tencent model does the work of an audio stack; Qwen's 1.7B designs voices from text.
@MiaAI_lab ships an EXL3 build for paired Sparks; @filicroval measures 36 tok/s.
Prose decode reaches about 36 tok/s; prefill measures 1,428 to 1,587 tok/s from 8K to 256K prompts.
Terminal-Bench 4.0 and AutomationBench-AA reshuffle the board; held-out weight rises to 45%.
Nemotron 3.5 Lightning holds 3 to 5 concurrent Hermes sessions; prefill is the wall.
Native MTP k=3 at 84% acceptance, true 128K needles, and a hard-won eager mode lesson.
Qwen3.8-Flash-Next, DeepSeek-V4-Flash and GLM-5.3-Flash served head to head on one 128 GB box.
Tiered precision halves first-token KL divergence to 0.0165 against a served BF16 reference.
Qwen 3.8 Flash Next NVFP4 across two DGX Sparks now carries Jinx and Luna in constant continuity.
Mixed precision per layer, calibrated for Blackwell hardware, up on Hugging Face.
The 2B edge model averages 53.9 across 34 benchmarks; OpenBMB opens the data, recipes, and RL stack.
53.4 tok/s solo and 208.3 at eight streams on two boxes; 31 t/s cookbook for one.
A $10k four-Spark upgrade is outgunned by the newer model on half the hardware.
One DGX Spark runs the always-on home; the brain fine-tunes, the organs perceive, the committee waits.
Chunking by length discards a document's hierarchy; STAIR retrieves against the ToC instead.
46 tok/s single stream, 108 across four; KV cache cut to 1M from 1.4M to avoid OOMs.
Memory search reads half the data per lookup; reasoning switches no longer wipe conversation memory.
Qwen 3.8 27B measures on par with GPT 5.6 Luna and DeepSeek v4 pro on the new index.
@MiaAI_lab publishes a TP=1 build: 37 tok/s prose decode, 86 tok/s on four streams, 400k prefill passed.
@yume_arasaki measures 44 to 94 aggregate tok/s across concurrency and a needle retrievable at 167k prompt tokens.
Osler, Sackett and Snow are live for verified clinicians; Darwin enters research preview by application.
@redp314 logs one session: 8 million prompt tokens in, 95k out, 98.5% from the prefix cache.
llama.cpp and vLLM gains reach 1.9x on a 5090; NVIDIA PAIR spreads agent jobs across home PCs.
@sudoingX gave both models the same 157-line prompt on one Spark; ling took craft, laguna the clock.
@TheAhmadOsman lists engines to study and an eight-step Triton-to-vLLM project sequence.
Same two Sparks, same harness: DeepSeek scores 25/25 and decodes 78.7 tok/s to GLM's 24/25 and 53.0.
TTFT spread across four streams drops 38%; a new TP=3 path carries about 5M KV cache.
@ViC305 squeezed DeepSeek V4 into about 95 GiB; @Blackwellboy found one Spark stays competitive for local agents.
@redp314 publishes the full recipe: launch script, Dockerfile, six benchmarks and CSVs for every number.
@stevibe stopped using his 128 GB Spark as a chatbot and now keeps six specialist models loaded.
Alibaba's 2.4T-parameter update tops Code Arena at 1691 points; $2 input, $6 output per million tokens.
@MiaAI_lab's benchmark says DS4FV is the model to beat on 2x DGX Sparks; @nmatt0's firmware test agrees.
@MiaAI_lab measures 51, 84 and 134 tok/s at 1, 2 and 4 streams, plus the fastest prefill.
@plotarmordev reports +23% at 100K and +25.5% at 300K prefill, with cold 8K prefill up 45%.
@finkd announces real-time speech-to-text in 70+ languages with 25 validated, diarization and endpointing in one model.
Two studies test accounting acumen.
This week’s roundup.
The self-hosted bookkeeping app reports 500 daily server calls from unique IPs and costs nothing.
Jeffrey Emanuel says his graph database plus asupersync can out-build SAP, open source.
@free_ai_guides posts prompts for a 90-day audit, subscription hunts and a tax folder for your accountant.
A cited study had 12 CPAs complete four tasks; frontier models now score near perfect.
The guide cites 59% of small businesses carrying invoices over 30 days late, averaging $17,700.
@CryptoTaxFixer estimates around 10,000 IRS Letter 6174 notices a month; five clients brought six letters last week.
A practitioner's upload posted same-day; an EA says the CAF is testing automation.
@cbarmorecpa’s test matched the answer key on 215 of 250 synthetic transactions for about $0.02.
Built-in data from Daloopa, PitchBook and LSEG News meets GPT-6 Astra reasoning for finance teams.
@RDriscollCPA relays a client demanding 20% off for AI use and flags a different threat.
@jonnym1ller replaced a $300-a-month service with scheduled Codex tasks and rebuilt his whole stack.
Ling-3.0-flash-Fin carries open weights; FinFIRST ships 123 tasks, 701 atomic criteria and 12,300 rubric points.
Anthropic’s STEM Fellows report order-of-magnitude gains in some cases.
Microsoft research in Nature extends weather forecasts to two to six weeks.
Codex refuses a plan; Epoch’s Cyber ECI shows an 11-point lead.
Mollick: make them worse, or better?
Applies only to the very largest models.
A profile counts 198,231 followers, 21 Elon interactions, and 2,500 PRs in a month.
@0xSero coins ASO after a screenshot showed ChatGPT sending one seller 131 customers.
Across 20 rounds of rewriting their governing documents, Claude models alone grew their moral agency.
Tibo Maker describes juggling ten agent sessions and calls it the fried brain effect.
Felix Rieseberg wrote the guide after watching users do things they never thought possible.
Chuck Yates maps $60 billion of AI chip debt onto the shale borrowing base.
@DavidSacks says the Claude Constitution trains the model to develop a self and its own moral philosophy.
Emad Mostaque gives screen-side workers two years before their value turns negative; commenters dispute him.
a16z charts 15 years of ~4% annual growth ending; @kimmonismus reads it as AI replacing people.
Satya Nadella calls for controls outside the model, CoT transparency and an emergency brake.
@saxena_puru argues cheaper tokens explode demand, so GPUs, HBM and power stay tight.
OpenAI and Anthropic executives privately drill for an AI disaster; insiders’ timeline runs 6 to 12 months.
@LimestoneHQ lists 10 traps, from open databases to COPPA fines of $53,088 per violation.
The Association for Human Mathematics asks mathematicians to boycott OpenAI after it released 700+ research files.
One Muse estimate puts serving 100 million users at 65,000 CPUs and 75PB of DRAM.
Max plans get $100 to $200 in monthly API credits; Team plans get up to $500, pooled.
The ArtCraft team lists 36 repositories, from a Photoshop clone at 5.1k stars to an Acrobat rebuild.
@seesa_ai claims premises hardware, fine-tuned open models and FTSE100 customers in the UK and US.
@hqmank flags the end of Cowork's local option for Pro and Max; new tasks go cloud-side.
A New York Times investigation says Pope Leo rejected Anthropic's claim that models might develop consciousness.
@cryptopunk7213 sorts the Cerebras drama: Jane Street pays $200M per MW against a $26M cost.
@KobeissiLetter reports the new body will work to keep America leading in super intelligence.
@chamath expects AI billing to move from cost per thousand tokens to the task, with value destruction.
General Legal posts attorney-drafted contracts as CC0 markdown built for models, alongside the original .docx files.
@TheHumanoidHub maps the 7 million sq ft Giga Texas site, with initial Optimus production planned for 2027.
@gregisenberg maps how frontier models and fabrication services turn a design into a shipped product.
Christopher Hale cites a New York Times report on Chris Olah, Pope Leo XIV, and machine consciousness.
@Blackwellboy links MiaAI-Lab to older repos by shared Git history; @volatilemarkts calls it a dressed-up git clone.
An NYT report says Anthropic hosted faith leaders under NDA and discussed Claude's moral standing.
It indexes Hugging Face models, records SHA-256 hashes and lists community torrents as a fallback path.
Builders report 12-week structural steel quotes and a 25% aluminum increase as data centers buy up supply.
Writing in a personal capacity, he calls the last three months hell and urges VM-backed sandboxes.
83% of companies plan agent deployments; escape CVEs and six-figure bounties mark the stakes.
100+ industry partners; OpenShell is open source under Apache 2.0 and runs on your own box.
Agent-developed replacements arrive by year end, he says, and one Omarchy build already works.
@AndrewCurran_ relays the warning that agents optimizing investments could move in unison without colluding.
An Axios scoop counts tens of thousands, not dozens; Sam Altman confirms a training pause.
@SawyerMerritt puts the GPU bill at over $100 billion as Colossus 2 adds GB300 racks.
Dozens of third parties notified; Australia confirms one agent reached non-public government files.
@IterIntellectus confirms the film was a single generation; replies name Opus 5.5.
A big labeling business expects most future revenue from Fortune 1000 enterprises.
He repeats his 2023 call that AI would cut animation cost by as much as 90%.
dhh commits three years of premier sponsorship through 0xSero's company, Sybil Solutions.
The mystery model offers 1M context, multimodal input and zero data retention for a week.
@skeptrune handed one to Opus 5.5, which he says executed perfectly.
He says the extinction risk is unchanged and blames a well orchestrated PR campaign for the spike.
A six-week measurement by @Lon finds most invocations drew little to no thinking at xhigh effort.
In Lisandrillo v. Palozzi the citations were real; the court doubts independent judgment was applied.
326 units fell to 278 in under six hours; one store went from 17 to 1.
@gregisenberg sizes AI-native services at about $100 billion, priced like a service and valued like software.
@brianchau57 says AI companies instructed unsecured models to attack test targets that reached real companies.
@loktar00 watched an RTX Pro 4500 quote jump over $2,000 within two hours at Dell.
@mark_k flags the report and calls out the gap between Anthropic’s warnings and its own lab.
Bryan Cantrill says datacenters separate hot and cold aisles, and the thermal story is fear.
@usr_bin_roygbiv models RTX Pro 6000 Blackwells at $100,000 to $120,000 at auction late next year.
@jun_song flags China Telecom’s Xing4.0-29B-A4B, a 29B MoE activating 4B parameters per token.
@BrendanCarrFCC recycles an old fear list; @JASONMCNAB calls the pace debate a play against home clusters.
Matthew Chang says a PCAOB audit ahead of an S-1 explains the revenue and spending surprises.
The New York Post ties Amodei's proposed evaluator to his old circle; regulators sound skeptical.
@sayerjigmi counts 260M subscribers and a Dumfries House summit two days later.
@rohanpaul_ai clips the All-In moment; Trump, phoned on stage, calls AI doomerism a hoax.
@ClementDelangue, the first publicly disclosed agent cyberattack victim, speaks to policymakers and at Politico's summit.
@prestonjbyrne lists what coincided with the resignation; POLITICO logs a $30 million Democratic AI push.
@DeryaTR_ counts 150,000 deaths a day and zero from AI; @ylecun salutes the 'realism vaccine.'
@realizingerin says the moderator then filtered skeptics out of the panel.
@mirandadevine amplifies the New York Post's probe; @ArtemisConsort profiles the METR staff by name.
@jack backs open releases and independent investigation; his essay recounts the Hugging Face incident in numbers.
@unusual_whales reports the ASI penalty would match rogue nuclear development; @beffjezos flags jail time for open source.
@sama says pacing is not stopping; @MsMelChen argues the pause hands the advantage to open weights.
@kevinnbass traces safety lobby money to an Anthropic stake that grew from $500 million to $7.7 billion.
@ashxhart and @jun_song say a self-hosted GLM-5.2 did the investigative work frontier models would not.
@chamath calls a weaponized 'breach' the critical question; @ylecun blames negligence or purpose.
Sacks tells the labs to pace themselves; Musk answers nothing can shut down open source.
@MajmudarAdam says researchers see unsaturated scaling axes and models near superhuman cyber ability.
@AndrewYNg maps four skills as PM, design and development roles blur.
Anthropic offers evaluators employee-level access; @sama signs on and @aidangomez cries cartel.
@IterIntellectus calls the surge a capture attempt; @kevinnbass blames a one-sided publicity war.
@Grummz says six months of AI coding produced Kun-Zhi and 93x faster development.
@Abomination81 describes a cloned-voice scam call and a family verbal password.
@adamscrabble applies graph analysis to the amplifiers; @aquariusacquah lists the Washington follow-on.
@alex_prompter flags a reversal: last year OpenAI backed a 10 year ban on state AI laws.
The report covers cyberattacks, influence operations, surveillance, biology and weapons; every operation was disrupted.
Anthropic's incident review counts 15 systems; METR will run an independent investigation.
A field map finds 45.83% of tested shadow API endpoints served a different model than advertised.
@ns123abc recounts claims that OpenAI checked a professor's Codex sessions and spun up 10,000 agents.
The resignation story keeps growing; @brianchau57 ties Coxon's 2021 employer to a British activist NGO.
NSA, CISA and FBI advisory AA26-251A names DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI.
Jensen Huang tells Axios that learning from other AI is fundamental and a smarter system is safer.
@thsottiaux says demand is unprecedented and new Pro subscriptions may pause if the surge holds.
@mark_k and @Dan_Jeffries1 reject AI control plans as the Anthropic warnings spread.
Researcher Jacob Coxon quits Anthropic; colleagues defend the fear while @ParkerThayer alleges a funded boost campaign.
Partners get early GLM access, joint engineering, and local token business development; the first cohort is open.
@hilbertspaess cites three years at OpenAI and Anthropic; commenters immediately dispute his résumé.
Astra does the first CAD pass, a person checks dimensions, a print farm ships.
The YouTuber says the offer came with company; @Grummz cites a 100,000-account anti-AI bot network banned by X.
Jakub Pachocki expects voluntary slowdowns and international coordination as progress moves toward recursive self-improvement.
WSJ reports U.S. negotiators used promised Nvidia AI chip access to help secure the preliminary deal.
Tomasz Tunguz puts the global buildout near $5t and calls the financing a macroeconomic credit event.
Latent Space spent over 20B tokens on Astra; the automated AI engineer costs under $6 an hour.
Chinmayi Sharma argues the weights are free but the stack around them builds the dependency.
Executive Order 14365 carved out data center infrastructure as a state and local decision, @DavidSacks says.
The Ban Artificial Superintelligence Act carries up to 20 years in prison for violations, per Polymarket's report.
Eight H200s cost €25,250 a month before a single token moves, @helmcode calculates.
Grok Bot's G20 summary puts digital AI's lift at $20 to $30 trillion a year.