@Luke85953209@sfxnz Pray tell what model, quant, and harness are you using? Qwen3.8 FN vs 27b are very different and Qwen Flash next with Strata runs good quants well on fair gaming hardware. Use a harness with /goal and direct it to iterate and I have tasks that have run for days til I stopped them
@coltfeltes@sfxnz Not only slower, but QwenFN is better at CLI! I love GLM as a thinker but Deepseek delivers outputs better than anything else Ive tried on 512gb vram, Qwen fills gaps and drives my operations on half the hardware and agentic tasks, so I find it hard to work GLM into the picture.
@Austen Chat is different from background agent work. I used to say the same. If you have multiple models running you can talk to a slower big brain planner if you have faster models doing the agent work but slow models doing agent calls becomes less dynamic long running background tasks
@MiaAI_lab Agreed! as somebody with 4 sparks already I'm in the gap now 4 vs 8... seems 6 requires a switch which is unavailable and 3/6 still seems to have reduced tp performance vs 2/4/8/etc. Would you do 6 sparks or M5Ultra which then could do rendering, embedding, training, etc faster?
@HermesWatcher I do put personality into the tts not in the llm generation. Some people hate on making different profiles for agents but I tie different profiles to different workflows and then control which profiles can communicate with one another. Is a built in workflow tool on the roadmap?
@thdxr Bro... who hurt you? lol I know where you're coming from. An industry led by people who not only don't understand the tech but don't care is frustrating! Simultaneously as a career convert 10years ago I think a passionate person being able to learn and build faster now is a win!
@MatthewBerman Is no one else worried we're talking so harshly over GPUs...? Am I the only one not ok with people going to jail for buying non-stolen GPUs? Do we really want this to be normalized? Do we want to have to register for GPUs at the store like we're buying weapons? Feels dystopian AF
@elshayib When the company isn't making a profit you cant make conclusions about their prices being better. When they are no longer burning multiple dollars for every dollar they make then we can celebrate
A brief history of AI Coding since Nov 2025:
Nov 2025: Opus 4.5 releases, OpenClaw gets built
Dec 2025: community agrees that models get good enough to handle 95% of coding tasks
Jan 2026: Ralph loops get popular, early efforts on loops/orchestration
Feb: Everyone feels knackered, too much strategic decision-making, fears around cognitive debt begin
Mar: Skills explosion as community converges around easy-to-share processes (superpowers already ahead of the game here)
Apr: most SOTA devs haven't written a line of code in months
May: "The unreasonable effectiveness of HTML" - more interest in richer review surfaces
Jun: "Loop Engineering" wins vaguepost of the year, signals more hype around process
Jul: Every talk in every conference becomes "here is my software factory"
Aug: Opus 5 speaks incomprehensibly, interest in ASD-STE100 peaks
Sep: Grok Bot gets popular, more interest in long-running "manager" threads which retain context and delegate work
Oct: The promised speed gains haven't materialised, "dark factory" feels a long way off still
Nov: No-slop manifesto releases, reading the code still appears essential
TL;DR: no-one is writing code anymore, and we are still figuring out what that means
@elshayib Big picture: 1)They are undercutting the USLab market so they cant raise prices. 2)All of this is debt spending for labs. None of them are profitable. The goal is Chinese subsidies hold out beyond provate equity. 3) For me the moat is self hosting so Kimi doesnt exist to me lol
@Tech2Wild I know some of my favorite techies I follow swear GLM is better. I think Qwen3.8FN is better at the command line and GLM is better at making us feel it is thinking. I can't give an example in my work where GLM did something qwen couldn't. GLM made me feel more confident though.
@elshayib Seems like there's always alternating rounds of American then Chinese model updates... I have to test before using them so I'm not rushing for new models right now lol. We have great prem weights that I'm still trying to maximize my use of.
Last week a momentous American AI milestone, the release of Inkling 2 Small's open weights by Thinking Machines Lab, was overshadowed.
For the first time in a long while, we have an American open-weight model competing in the same performance tier as some of China's leading open-weight models, while holding its own against entry-level frontier offerings, as shown in the attached benchmark.
I have reservations about how Artificial Analysis recently recalculated its Intelligence Index, but the comparison is still remarkable. Inkling 2 Small scores alongside Gemini 3.8 Flash, above Luna GPT 5, and within reach of Grok 4.7 and Claude Haiku 5.5.
Jensen Huang recognized the potential of Mira Murati's Thinking Machines Lab and invested in its contribution to the broader AI ecosystem. He's always been a fierce competitor, but what sets his approach apart is his willingness to strengthen the ecosystem around NVIDIA rather than simply protect his position by limiting competition.
Some worried NVIDIA's involvement might undermine Thinking Machines' independence. Yesterday's open-weight release offers an encouraging counterpoint.
Many compete by trying to sink everyone else's ships. Jensen labored to be the tide that lifted the entire industry. Sadly, I suppose all tides fall eventually, too. Perhaps the tide will return...
Don't block normal system admin. Tune to stop adding extra security rules I never chose in my agent docs that cause other agents to stop making the progress they were known for at one point. Provide better alternative credential brokering and sec options instead of just blocking.
I'm not doing extra sign ups changing my privacy rights to get my work done, I'm just switching providers.
@April_Wangtong@thsottiaux Agreed! Sol has clearly been tuned to stop more. It used to be a "Rottweiler" as Theo says but now I'm begging it to do anything. Even with goals set it looks for reasons to stop. This Chat/work/ codex is still messy. I dont get it and if codex != work then make clear why.
@thsottiaux Treat the $20 users (who built the paid user base) and the $8/API paid with the full features of the platform (Sol, Terra or Luna for cheap Dot for $20/$8 is fine). Lower quotas already limit use. Make sure imposed security controls don't block normal IT work. Make a $50 plan.
@liuliu On my M5Max studio Qwen 3.8 NVFP4 in Splash over 100t/s. M5Max MBP Qwen 3.8 Flash Next EXL3 4bpw at 60-100t/s.
Tested and they score near fp8 on terminal bench and deep swe benchmarks with similar token output per task meaning near lossless IMO.
@theo T3Code doesnt welcome local and you clearly dont value open weight contributions to the ecosystem so you are just falling out of that conversation. There's a ton of energy in self hosted open weights right now. Claude is best but also being free from frontier lab limits is great!