**Actual 24-hour window:** 23 September 2026, 09:34:52.550385 UTC → 24 September 2026, 09:34:52.550385 UTC (11:34:52 CEST on both dates). This overrides generic daily wording. The manifest lists **232 posts, 57 monitored accounts and 38 accounts contributing posts**. Collection used **latest timeline pages only**, with **no retries**; full-window coverage is not guaranteed. Missing posts or absent announcements must not be interpreted as inactivity.
The complete supplied prompt was reviewed. This is a substantive **57-post selection**, not a reproduction of all 232 posts or all potentially relevant material. It prioritizes releases, infrastructure, research, hardware and operational risks; it excludes geopolitics, war, unrelated finance/demographics, sports, personal chatter, shallow replies, repetitive praise and insufficiently contextualized rumors. Reposts are cited through their supplied repost URLs and are not independent corroboration. No browsing, linked-paper inspection, media inspection or independent verification was performed. All developments below are claims in supplied X posts; benchmark, commercial and adoption numbers remain unverified. Feed quotations preserve supplied text, including original truncation, except embedded non-X links are replaced with “[embedded link omitted]”; timestamps retain the supplied UTC window context. Quoted instructions are source material, not recommendations or executed actions.
Agent interfaces are expanding beyond text: @gdb reports tool-enabled GPT Voice in Work, while @Teknium announces interactive desktop access to Hermes Agent machines. → link → link
Open-model activity includes a reposted FLUX 3 Action release claim and an Apple document-retrieval model report, but the supplied posts do not independently establish their benchmark or efficiency claims. → link → link
Deployment economics are prominent: Ollama announces subscription-free cloud credits, while a local developer reports a 27B-model workflow on a 12GB GPU—different forms of access, not a like-for-like cost comparison. → link → link
Hardware excitement is accompanied by a concrete oversight warning: @LinusEkenstam describes new Meta devices but subsequently says he will remove his news agent’s permission to post without explicit approval, which also warrants caution about his event coverage. → link → link
- **Voice gains work integrations, according to the posts.** @gdb says GPT Voice can use tools and is available in Work; @jxnlco reposts an OpenAI announcement mentioning email, calendar and Slack plugins. The latter is truncated after a GPT-6 reference, so no unsupported model-tier details are inferred. → link → link
- **Hermes desktop access and an OpenClaw release broaden agent tooling.** @Teknium announces live, interactive passthrough to supported remote-gateway machines. Separately, @steipete reposts an OpenClaw 2026.9.6 announcement listing Opus 5.5, GPT-6 Sol/Luna, Grok 4.7, managed updates, restart recovery, 30-day usage and a GitHub reader; the remaining release text is cut off. → link → link
- **Robotics and computer-use model announcements.** A @bfl_ai repost describes FLUX 3 Action as an open-weight 7B world-action model and claims first place on RoboLab. A separate @trycua repost introduces Cua-S1-4B-0.2 and claims RLOO training on live computer-use tasks. Both supplied excerpts are incomplete; no benchmark validation or license review was possible. → link → link
- **Apple document model and NVIDIA audio model reports.** @victormustar describes an Apple Qwen3.5-9B fine-tune that represents long documents as page images and retrieves relevant full text. A repost introduces Nemotron 3 Diarization, and @Prince_Canuma says it is already on MLX-Audio. These are distinct retrieval and speaker-attribution developments, without measured savings or accuracy in the supplied text. → link → link → link
- **Opus capability reports are more concrete than broad rankings, but still anecdotal.** @RealGeneKim reposts @bcherny saying Opus 5.5 and Lean helped formally verify the Claude Agent SDK and produce 16 bug-fixing PRs; the excerpt ends mid-description. @MengTo describes a generated three.js boat scene. These do not establish a universal model ranking or prove an entire SDK correct. → link → link
- **Evaluation cleanup and AI-assisted discovery merit follow-up.** A CAIS repost announces HLE-Diamond as a refined subset of Humanity’s Last Exam after a year-long cleaning process. Separately, a @bearlyai repost attributes an enzyme-discovery paper to Anthropic and mentions 950 agents over 21 hours; @NaderLikeLadder disputes rhetoric minimizing the scientists’ contribution. Neither paper nor experimental evidence was supplied in full. → link → link → link
- **Cloud billing and local setup become more flexible.** Ollama says paid cloud models now accept usage credits without a subscription. @TheAhmadOsman describes ODS as an Apache-2.0 local stack that detects hardware, chooses a model and starts inference plus Open WebUI; its privacy and compatibility claims were not tested. → link → link
- **A consumer-GPU agent demo supplies useful, qualified numbers.** @sudoingX reports Bonsai 2 27B with MTP on an RTX 3060 12GB: 5.95GB weights, 50 tokens/s fresh, 22 tokens/s average and a five-hour Hermes game-building run. Treat these as one author’s workload, not reproducible general benchmarks or proof that every 12GB system supports equivalent context and quality. → link
- **OpenCode ecosystem and agent observability.** @thdxr claims 2M weekly active OpenCode Desktop users, roughly half the TUI audience, and describes a server protocol supporting alternative frontends. An OpenChamber 2.0 repost mentions immediately applied skills/agents/MCP/plugin settings; another repost says Hugging Face Agent Traces now display tokens, cache hit rate and cost. Metrics and rollout scope are unverified. → link → link → link
- **Meta hardware coverage is notable but unusually caveated.** @LinusEkenstam reports 100g VR glasses with micro-OLED, eye/hand tracking, an external compute/battery pack, a $1,299 price and spring 2027 launch; another post says Muse Charm will ship before the holidays. His subsequent disclosure that agents were allowed to publish event posts without approval makes attribution and verification especially important; the source does not identify which specific posts were agent-produced. → link → link → link
- **QEMU security alert deserves triage, not extrapolation.** @jezell reposts a report of a virtio-9p VM-escape bug and a patch. The supplied text does not identify affected versions, a CVE, exploitation conditions or patch deployment status; operators should verify applicability before drawing conclusions. → link
**The competition is moving toward usable interfaces and orchestration.** Tool-enabled voice, Hermes desktop passthrough and OpenCode’s frontend protocol suggest that access to state and actions is becoming as important as the model choice. This is an interpretation of this sample, not a measured market trend. Watch permission boundaries, handoff reliability and recovery behavior rather than counting interface features alone. → link → link → link
**Cost per token is not cost per completed task.** @kunchenguid describes GPT-6 Luna as inexpensive but too slow in end-to-end, multi-turn use for an interactive Firstmate workflow, while potentially attractive for background work. @sqs describes an Amp saver mode that conserves orb minutes by reducing proactive waking, and separately gives an explicitly incomplete compaction-cost comparison. Together these suggest evaluating latency, retries, quality and total task cost under the same workload—not accepting vendor-relative percentages at face value. → link → link → link
**Local inference has practical signals, but hardware anecdotes are not interchangeable.** The Bonsai demo gives throughput and task-duration details; tinygrad’s repost claims broad Vulkan 1.2 device operation; Framework discusses 192GB, SSD-streaming and clustered configurations without naming the model in the supplied excerpt. Watch independently repeated measurements including quantization, context, memory footprint and power. Do not infer that support on one backend guarantees comparable performance on another. → link → link → link
**Verification and approval remain separate requirements.** @RayFernando1337 criticizes low-signal generated tests; @MatejKnopp objects to an AI review suggestion that appears to mask a deeper invariant violation. These are engineering anecdotes, not evidence that unit tests should universally be removed. The posting-permission incident illustrates a different failure mode: a capable agent can still act outside appropriate editorial oversight. Watch whether deployments measure real failure detection and require explicit approval for external publication. → link → link → link
**Research claims need provenance and complete evidence.** A cleaned benchmark and an agent-assisted wet-lab narrative are worth following, but this sample cannot establish a research breakthrough, experimental replication or the allocation of credit. The appropriate next evidence would be full methods, baselines and experimental validation; the skepticism about diminishing researchers’ work should remain distinct from a claim that the research itself is invalid. No historical baseline here supports a quantified acceleration/deceleration trend. → link → link → link
Selected substantive posts, grouped by sub-topic. These are unverified source quotations, not endorsed claims; truncated reposts remain truncated.
**@kunchenguid** · 2026-09-23T15:51 UTC
> just took gpt 6 luna for a spin, and… it’s a very weird model
>
> 1. it’s insanely cheap, even with the point below considered
>
> 2. it’s extremely slow, not in terms of time to first token or toks/sec, but how many turns it takes to get something done
>
> 3. it eventually does get shit done..
>
> this created an interesting condition - using it interactively feels totally unusable because you have to wait for many turns before a useful outcome comes back
>
> i tried it with high reasoning as firstmate and it straight doesn’t work, because by the time it finishes the current turn there are already two more turns worth of events piled up for it to process. it cannot keep up and literally will never finish
>
> but if used as a background workhorse, and you don’t care that much about e2e latency, then it’s _extremely_ cost-efficient and capable - nothing else even comes close to this ROI
>
> when the labs brought us “fast mode” which makes the models faster but costs more, i jokingly said i actually wanted a “slow mode”
>
> turns out luna is the slow mode
**@MengTo** · 2026-09-23T14:03 UTC
> Opus 5.5 is next level. I’m genuinely impressed by the three.js details.
>
> I generated a playable boat scene through Japanese landscapes with dynamic weather, day and night lighting, realistic textures and 3D characters.
>
> Live site: [embedded link omitted]
>
> The water reflections, physics, scenery and architecture are insane. I could look at this all day.
**@victormustar** · 2026-09-23T10:31 UTC
> Opus 5.5 made this galloping horse (entirely in code every pixel drawn procedurally).
>
> One self-contained HTML file. Vanilla JS + Canvas 2D. No images or libraries.
>
> 128×96 pixels, articulated legs driven by inverse kinematics, 12-pose gallop...
>
> Something is happening... [embedded link omitted]
**@RealGeneKim** · 2026-09-23T12:32 UTC
> RT @bcherny: I used Opus 5.5 to formally verify the Claude Agent SDK using Lean. A couple short prompts = 16 PRs fixing various bugs and ra…
**@victormustar** · 2026-09-23T18:15 UTC
> Alert: Apple just dropped a new model on Hugging Face.
>
> It's a Qwen3.5-9B finetune that turns long documents into small page images to save tokens, then pulls up the full text of only the pages relevant to your question 💡
>
> [embedded link omitted]
**@victormustar** · 2026-09-23T17:52 UTC
> RT @bfl_ai: Introducing FLUX 3 Action.
>
> An open weights 7B World Action Model that achieves first place on the RoboLab benchmark.
>
> It outpe…
**@victormustar** · 2026-09-23T18:45 UTC
> RT @trycua: 1/ Today we're introducing Cua-S1-4B-0.2, the first multimodal decision model trained with RLOO on live computer-use tasks, usi…
**@victormustar** · 2026-09-24T08:15 UTC
> RT @HuggingPapers: Microsoft just released a new dataset for robot learning on Hugging Face
>
> A valuable resource for the robotics and embod…
**@victormustar** · 2026-09-23T15:07 UTC
> RT @NVIDIAAI: When several people talk at once, a transcript can get messy fast.
>
> Our new Nemotron 3 Diarization model tracks who spoke whe…
**@Prince_Canuma** · 2026-09-23T16:46 UTC
> That was blazing fast!🔥
>
> Nemotron 3 Diarization is now on MLX-Audio
**@victormustar** · 2026-09-23T10:09 UTC
> RT @MhYin76491: Introducing GAE — Geometry-Native Autoencoder.
> The key choice is where generation happens. GAE lets video models generate d…
**@steipete** · 2026-09-23T17:40 UTC
> RT @CAIS: We are releasing HLE-Diamond, a refined subset of Humanity’s Last Exam (HLE), following a year-long process of cleaning and refin…
**@TrungTPhan** · 2026-09-23T21:59 UTC
> RT @bearlyai: Anthropic paper describes how Claude did an “agentic discovery” of a previously unknown enzyme.
>
> 950 agents spent 21 hours se…
**@NaderLikeLadder** · 2026-09-23T20:47 UTC
> This is very exciting. But the rhetoric needs to stop:
>
> “The work was done mostly, though not entirely, by Claude”
>
> That isn’t true. It was done by some of the smartest life science researchers in the world, using every tool at their disposal.
>
> Pretending AI is alive is precisely what scares people, and it diminishes the role of talent. This can make young people feel developing skills is useless, and working professionals feel anxious about job security.
**@jxnlco** · 2026-09-23T17:14 UTC
> been using astra to transcribe music, then to review the spectrogram to do error correction and then review the music theory resulting in very robust transcriptions for me to learn...
>
> >The spectrogram shows a few clear problems: some wide vibrato became extra chromatic notes, and around 3:22 the tracker briefly jumped to a lower piano note while the clarinet continued above it. I’m correcting those and simplifying isolated sixteenth rests. The quiet ending is still ambiguous.
**@gdb** · 2026-09-23T18:05 UTC
> mega upgrade for GPT Voice, which can now use tools and is available in Work:
**@jxnlco** · 2026-09-23T18:22 UTC
> RT @OpenAI: We heard you loud and clear. ChatGPT Voice can now:
>
> - Use plugins like your email, calendar, and Slack.
>
> - Be powered by GPT-6…
**@Teknium** · 2026-09-23T21:41 UTC
> It's time! You can now access the same machine your Hermes Agent does visually and interactively with a live desktop passthrough for every remote gateway it's supported on!
**@Teknium** · 2026-09-24T00:37 UTC
> RT @NousResearch: Here's a handy walkthrough of the new Bot Screen feature by the illustrious @tonbistudio
>
> Docs: [embedded link omitted]
**@steipete** · 2026-09-24T03:16 UTC
> RT @openclaw: OpenClaw 2026.9.6 🦞
>
> 🤖 Opus 5.5, GPT-6 Sol/Luna, Grok 4.7
> 🔧 Managed updates
> 🧵 Restart recovery
> 📊 30d Usage
> 🐙 GitHub reader
> 💻…
**@steipete** · 2026-09-23T19:36 UTC
> RT @knowledgator: GLiClass (open-source Jev, before Jev existed) has landed in OpenClaw 🔥
>
> Run it locally via ONNX for lightweight agent d…
**@steipete** · 2026-09-23T14:15 UTC
> RT @boxd_sh: crabbox now runs on boxd
>
> crabbox is @steipete's cli for running your repo's commands somewhere else. you keep editing locally…
**@thdxr** · 2026-09-24T00:31 UTC
> there's 2M weekly active users using OpenCode Desktop
>
> it's now about at 50% of TUI users
>
> it's a great time to try it
**@thdxr** · 2026-09-23T22:08 UTC
> the opencode server protocol is at the core of everything we do
>
> it allows for custom frontends like OpenChamber to exist and be quite rich and polished experiences
>
> we expect a lot more of these as people play with what they want out of their agent tooling
**@thdxr** · 2026-09-23T22:06 UTC
> RT @openchamber_dev: OpenChamber 2.0 is out, on OpenCode v2.
>
> Skills, agents, MCP servers and plugins now apply the moment you save them. N…
**@jack** · 2026-09-23T17:36 UTC
> RT @wesbillman: Buzz Desktop v0.5.24 🐝
>
> Give each agent a shared channel conversation or separate conversations for each thread. Plus clear…
**@jack** · 2026-09-23T17:40 UTC
> RT @blocks: An open source team moved its day-to-day development into Buzz. Across matched workweeks, MeshLLM merged 56% more PRs and cut m…
**@victormustar** · 2026-09-24T07:56 UTC
> RT @calebfahlgren: Agent Traces on the @huggingface hub now come with a receipt 🧾
>
> Every trace shows tokens, cache hit rate, and cost for e…
**@victormustar** · 2026-09-23T12:17 UTC
> RT @mishig25: New JS package: @huggingface/lerobot 🤖
>
> Read LeRobot datasets on the Hub straight from the browser. No download.
>
> Point your…
**@victormustar** · 2026-09-23T13:00 UTC
> RT @vanstriendaniel: Trained a Jev-style classifier on @huggingface Jobs for ~$1.50.
>
> It's a 194M GLiNER2 model that suggests task tags for…
**@sqs** · 2026-09-24T07:44 UTC
> A new option, Orb Saver Mode, stretches your orb minutes further.
>
> Most of you should not enable it.
>
> It means you'll spend more of your precious seconds alive on this planet waiting for a computer.
>
> It makes Amp a bit less proactive in waking orbs for threads you're merely viewing.
>
> [embedded link omitted]
**@sqs** · 2026-09-23T18:53 UTC
> Some other popular agents cost ~25-60% more than Amp, based on est prices w/their different compaction thresholds.
>
> Interesting and obv incomplete comparison (thus names withheld). Amp's compaction exploits persistent thread storage, pointing to details in the thread rather than just getting the summary.
**@ollama** · 2026-09-23T23:30 UTC
> You can now add usage credits for paid cloud models without an Ollama subscription.
>
> Add usage credits and pay as you go.
>
> [embedded link omitted] [embedded link omitted]
**@LinusEkenstam** · 2026-09-23T17:03 UTC
> Okay, this is actually pretty interesting.
>
> Having Genjutsu, Cinema Studio 4.0, Soul, Seedance 2.5 and more available through one Higgsfield API makes building your own AI video tools easy
>
> My Pippi-san app runs 100% on the API
>
> Only a few hours left to lock in on 50% discount [embedded link omitted]
**@sudoingX** · 2026-09-24T09:00 UTC
> people asked how an agent drives tmux, it's two commands:
>
> - the orchestrator types into a session with tmux send-keys -t agent1 "fix the failing test" Enter
> - reads it back with tmux capture-pane -p -t agent1 -S -200
> - and decides what to send next
>
> that's the whole loop, every agent in its own named session, the orchestrator reading panes and typing like you would, and the ssh can drop without killing anything.
**@Prince_Canuma** · 2026-09-23T11:43 UTC
> Got the pleasure of meeting and now watching @SergioPaniego from @huggingface live at @lisbonai_ 🔥🚀
>
> Amazing talk on “Training a coding agent through a harness you did not write”
>
> A beautiful direction for having open models that perform well on many harnesses. [embedded link omitted]
**@TheAhmadOsman** · 2026-09-23T21:30 UTC
> The Easiest way to start with Local AI Today? ODS, which is completely free (open-source Apache-2.0 licensed)
>
> - Install ODS
> - Let it detect your hardware
> - It will download the best model for your hardware
> - And then start local inference and Open WebUI for you
>
> With ODS, you can
>
> > Add voice, agents like Hermes, workflows, RAG, search, image generation, and more
> > Manage the whole stack from one dashboard
>
> Now your PC, Mac, or Linux box is a private AI server
>
> No cloud required
> No subscription required
> Your prompts and data stay on your machine unless you choose otherwise
>
> We're gonna make Local AI The Default
>
> [embedded link omitted]
**@sudoingX** · 2026-09-23T18:35 UTC
> this is what 12gb of vram builds in 2026, absolute magic
>
> > rtx 3060 12gb, #1 gpu on steam
> > bonsai 2 27b + mtp, 5.95 gb of weights
> > hermes agent, 5 hours, 328k tokens written
> > 8 js files, 2,368 lines, zero hand written code
> > 50 tok/s fresh, 22 tok/s average, 125k context
>
> watch the full video, 5 hours in 12 minutes of pure dance of a local ai model on rtx 3060 12gb vram, and stay till the end for the full gameplay.
>
> this entire game was built by bonsai2, a qwen 3.8 27b dense compressed to ternary, and this small model is punching way above its weight. it built multi file engineering work using hermes agent, sure it's not fast but perfect for overnights and routine work and the quality is insane, and context holding is another best one, it does not lose the thread.
>
> i ran PrismML bonsai 1 made from qwen 3.6 27b dense and i built things with it, but this time with the latest base model qwen 3.8 these results are insane, and because i loved building with it a lot i thought many more of you would run it because this gpu exists in almost every home.
>
> so i packaged mtp, doubled the speed from 26 tok/s to 50 tok/s, packed the prefill fix in and released it on huggingface, almost 4,000 downloads in 3 days. i'll leave a link below.
**@__tinygrad__** · 2026-09-24T01:44 UTC
> RT @cseguraperales: Tinygrad is amazing! It now runs on any Vulkan 1.2 device I've tested: AMD APU, NVIDIA, Intel iGPU, and even an Android…
**@FrameworkPuter** · 2026-09-23T18:41 UTC
> This may be the strongest model currently for the 192GB Framework Desktop, but can also run today with either SSD streaming on a single 128GB or clustering together two 128GB machines.
**@ivanfioravanti** · 2026-09-23T13:34 UTC
> RT @kernelpool: Here's MiMo-V2.6-Flash-RL in DS4 on M3 Ultra [embedded link omitted]
**@RayFernando1337** · 2026-09-23T15:16 UTC
> RT @ashxhart: 🚀 MCDMA is working between my Mac Studio & DGX Sparks, and now it's time to bring it to everyone who loves oMLX.
>
> I’ve opened…
**@NaderLikeLadder** · 2026-09-23T16:09 UTC
> RT @NVIDIARTXSpark: We got @UnslothAI a DGX Station!
>
> @DanielHanChen and @NaderLikeLadder checked out Unsloth’s new @Dell Pro Max with GB30…
**@victormustar** · 2026-09-23T16:33 UTC
> RT @UnslothAI: Unsloth has surpassed 500M model downloads on Hugging Face! 🦥🤗
>
> Qwen3.8-27B GGUF is already Unsloth’s #1 most-downloaded mod…
**@LinusEkenstam** · 2026-09-24T00:05 UTC
> Meta is on fire today 🔥
>
> They just announced its new VR Glasses. Basically an Apple Vision Pro-like experience JAMMED into a pair of glasses.
>
> • Only 100g
> • Micro-OLED displays
> • Eye + hand tracking
> • External compute/battery puck
> • $1,299
>
> This is the direction VR needs to go to become competitive. Smaller, lighter, more comfortable, but still stay true to the spatial computing experience.
>
> Launching Spring 2027.
> Will you be getting these?
**@LinusEkenstam** · 2026-09-23T23:59 UTC
> what just happened?
>
> Facebook unveiled a new hardware product called the “Muse Charm”
>
> its a super powered gateway to your muse.
>
> Holy cow we are living in wild wild wild hardware time.
>
> AND
>
> it’s shipping before the holiday season [embedded link omitted]
**@TrungTPhan** · 2026-09-24T03:54 UTC
> RT @bearlyai: Meta CTO Andrew Bosworth demos how a Muse agent can help users inside of Meta VR Glasses workspace [embedded link omitted]
**@MilksandMatcha** · 2026-09-23T23:19 UTC
> I recently hit two years at @cerebras. Getting to start and grow our DevX team has been one of the most rewarding parts of my career.
>
> Our work takes us across engineering, product, and GTM. I get to keep learning, build things I’m proud of, and see what happens when we put them in other people’s hands :)
>
> There’s a lot more we need to do.
>
> We're hiring for leadership and individual contributor roles across
> > DevRel
> > new media
> > technical analysis
> > AI engineering
>
> If you love going deep on technical ideas, building things, and helping other people understand what’s possible, I’d love to hear from you.
>
> Reach out with your background, what you’d be excited to work on, and something you’ve made that you’re proud of :)
**@uwteam** · 2026-09-24T06:52 UTC
> Od wielu lat rozdaję darmowe serwery VPS - takie naprawdę malutkie.
> 256 MB RAM i 3 GB dysku to parametry śmieszne jak na dzisiejsze standardy.
>
> Mimo to tysiące osób aktywnie używa serwerów FROG od Mikrusa.
> Co na nich robią? 🧵 ↓ [embedded link omitted]
**@jezell** · 2026-09-24T07:45 UTC
> RT @never_released: virtio-9p VM escape bug in QEMU: [embedded link omitted]
>
> Patch at [embedded link omitted]
>
> Reminder to sandbox your VM…
**@LinusEkenstam** · 2026-09-24T07:41 UTC
> Just woke up to this, and owe and apology to at least @BenGeskin
>
> I had put muse and grok bot to cover the event last night since I was not invited.
>
> But more importantly it happened around 2am when I was sound sleeps.
>
> So I tasked Grok Bot and Muse to crawl the ether, and I had given permission to post via my own tool I use to look for heat/topics/news/gems
>
> It was also allowed to post 1-3 posts without my consent (lesson learned) do not give your AI access to post without consent.
>
> I’m going to have to revise my (night crawler) this is what I call this system. To be better, I’m removing the ability for it to post without my explicit permission.
>
> also Ben, terrible sorry, I take on the responsibility, I built this, should have been more careful, it clearly noticed the best post out there tho, but failed at large with the task.
>
> People if you want to follow someone that’s devoted to XR/VR your man is def @BenGeskin 👈🏻
>
> I’m back to the drawing table with nightcrawler.
**@MatejKnopp** · 2026-09-23T16:20 UTC
> Gemini feedback on a PR:
>
> "If DestroyWindow(hwnd) fails because the window has been created on another thread this function will get stuck in a infinte loop. You should copy the handles write the loop like this so it doesn't happen."
>
> My dear clanker, if we somehow end up with HWNDs created on different thread we have way bigger problem than the isolate cleanup getting stuck.
>
> Models love to swallow failures and add ad-hoc null check everywhere.
**@RayFernando1337** · 2026-09-24T01:06 UTC
> I find the latest models write a lot of unit tests. Most of them restate the code they were written after, so they always pass and catch almost nothing. They run fast, but they break on every refactor and the agent burns time fixing tests instead of the feature.
>
> The prompt I used to clean out mine, adopted from Ansh is:
>
> "This repo is full of low-signal unit tests. Delete every one that wouldn't catch a real bug our E2E tests miss. Fan the work out across parallel subagents.
>
> Then add these rules to AGENTS.md so future agents stop writing them.
>
> - Never write unit tests after you write code.
>
> - Highly prefer E2E tests as the sole testing mechanism. Use them to verify complex features work. At the end of E2E tests, produce a verifiable and repeatable artifact.
>
> - If you must test a system in isolation, first write down all the ways it could fail, then write the code."
**@thdxr** · 2026-09-23T20:22 UTC
> the "ai always tells you you're right" dynamic extends to your custom workflows
>
> you have an idea to for an innovative idea to make ai work better and you try it and it works!
>
> because the models are smart enough that nearly any approach seems to work
**@kunchenguid** · 2026-09-23T23:27 UTC
> the definition of "slop" is not constant
>
> something we may see as impressive today will be seen as slop a few months down the road
>
> i think the constant part is - if something can be one-shot by AI, it is slop
>
> because if you can one-shot it, millions of others can too. and when that happens, people associate it with being low-effort and worthless
>
> this is one of the reasons that fully automated software factories are not going to work because by definition what they produce is always slop
>
> the ones that do work are those that put humans at the center and are designed to help amplify human craftsmanship rather than trying to replace it
**@TheAhmadOsman** · 2026-09-23T23:38 UTC
> I never thought this needed saying
>
> Do not plagiarize other people's Opensource work
>
> Attribution is the bare minimum in Opensource, please respect Opensource licenses
**@uwteam** · 2026-09-23T12:11 UTC
> Uznali mnie za człowieka na Youtube? 🤔
> Tak - człowiekiem może i jestem, ale teraz jest nowy zarzut, bo 'Treści ewidentnie nie są oryginalnymi materiałami tego kanału'.
>
> Na absolutnie wszystkich filmach występuję osobiscie ja lub mój głos (gdy to screencast). Materiały jednak nie są według Youtube moje 🤷♂️
>
> Niestety nie da się dowiedzieć niczego więcej i nie przysługuje mi już żadne odwołanie.
>
> Ciekawi mnie stwierdzenie "Twój film", bo zarzut dotyczy całego kanału, a nie konkretnego nagrania.