stellz @piss.beauty · 9hit bangs btw less verbose reasoning (which was the primary annoying thing about DS to me) thus less time and tokens to accomplish the same task, also just better performance 2510 51
stellz @piss.beauty · 9hdetails direct from themapi-docs.deepseek.comDeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real | DeepSeek API Docs🚀 DeepSeek-V4.1-Flash-1017 is here! 061 6
stellz @piss.beauty · 22ha lowest tier review is like 25-30 seconds of gemini flash + a free openrouter model (nemotron 120b rn) so it's literally free.99 010 1
stellz @piss.beauty · 22ha review-low tier review (the standard 2nd round review tier) often costs less than a cent to run, like $0.006-$0.008, if the change isn't big. and like $0.02 for a really big one main (medium) tier is like %50 more that, would be more without Gemini subsidy but still like 5 cents 110 1
stellz @piss.beauty · 22ha lot of people would be surprised just how cheap this review process is with the right models because of freeriding on the gemini for ~8 more months I get half of the reviews free, but that's a bit < $20 sub equiv the rest is all api pricing for luna / haiku / ling 3.0-flash / rare sol or sonnet 130 3
stellz @piss.beauty · 22hshould I use this to update my blog stack and then blog about the approach. would anybody read 370 7
stellz @piss.beauty · 22hI'm already doing it with free nemotron 120b on openrouter, it's a bottom-tier reviewer lol 130 3
stellz @piss.beauty · 17/10/2026have I crossed the threshold into full ai psychies now by showing y'all this 010 1
stellz @piss.beauty · 17/10/2026really useful for optimizing whatever processes you're running in the loop over time the agent assesses quality of the reviews when it writes the ledger entries, and it does work to validate the review findings as part of its natural workflow anyway, so the scores it assigns aren't out of nowhere 100 1
stellz @piss.beauty · 17/10/2026@jkachmar.com i remember you asked me to tell you how i ended up customizing pi i only went slightly sicko modetangled.orgpiss.beauty/pi-rigmy pi harness customizations, including a structured, multi-model system of iterative work 2420 42
stellz @piss.beauty · 16/10/2026there's even a reviewer calibration mode where you launch a panel of models to review the same diff, & then the main agent validates their findings before doing a quality comparison between the reviews it got here i'm comparing gemini-medium, luna-medium, luna-high, & ling-high 140 4
stellz @piss.beauty · 16/10/2026they even squash out these intermediate commits from the history before moving on from each step, since we mostly make the intermediate commits to easily generate a diff since last review, & only secondarily for checkpointing the work it's all really focused on continous mechanical deslopifying 140 4
stellz @piss.beauty · 16/10/2026the 2nd round of review for each step uses a weaker review panel than the 1st for cost reasons, but can get away with it because we ask them to focus on verifying that fixes applied after the previous round of review are good includes using a intermediate commit to supply a diff since last review 130 3
stellz @piss.beauty · 16/10/2026i don't really i juggle 2-3 projects throughout the day & let 1 main agent work in each at once each calls skill:review (or review-low, review-high, as needed by risk level) when it needs, that launches 2 independent reviews which main agent then validates the findings of + fixes, then repeat 260 6
stellz @piss.beauty · 16/10/2026agents know to use this iteration loop & i set them to auto mode, so i can put deepseek on it in the morning & it will churn for hours using this review cycle, resulting in a branch which is pretty much ready for PR even knows how to run benchmarks, so it will bail on an optimization if invalidated 020 2
stellz @piss.beauty · 16/10/2026it's very effective for producing clean git branches with logical commits, which can easily be reviewed by upstream maintainers each commit gets 2 two rounds of review with 2 reviewers each then there's a skill that does the same thing but considers the branch as a whole, makes sure it is PR-ready 2100 10
stellz @piss.beauty · 16/10/2026granted, i have flash model madness - i use a custom panel of different flash models that review each other's work in a structured way to do %95+ of work including tracking quality vs cost of review panel members over time 3191 19
Reposted by stellz∿ @somewhere.systems · 16/10/2026We've developed a new inference engine, called 'e'. It's the closest thing to alien technology which we have, and pushes 10,500 tok/s aggregate on a single DGX Station with the new @NVIDIAAI Nemotron 3.5 Super VL model. actual.inc/company/blog...actual.incActual Computer › Case Study : GB300 and Nemotron 3.5 Super VLDepartment-scale AI from one NVIDIA DGX Station: Nemotron 3.5 Super VL at 8,000+ tokens per second, shared by your whole team through api.actual.inc. 91238 123
stellz @piss.beauty · 16/10/2026I haven't I love the sparkle. it sounds like it wasn't amazing? 240 4
stellz @piss.beauty · 16/10/2026these cases were like 2/3s the normal price for a short while so I was like sure I'll have as many $1.17 celsius cans as I can reasonably store 0140 14
stellz @piss.beauty · 16/10/2026that's just you knowing that coffee tastes good. this is caffeine addiction speaking 6570 57
stellz @piss.beauty · 16/10/2026my main critique is cern should get a better partner than mistral 040 4
stellz @piss.beauty · 16/10/2026counterpoint: attie may not be good but shameless investorbait sustains companies 140 4
stellz @piss.beauty · 16/10/2026which is different from the skating rink themed bar without skating that also has DJs 040 4
stellz @piss.beauty · 16/10/2026I think they're just cranking opus for flash model level problems 1100 10
stellz @piss.beauty · 15/10/2026they had a 4hr 53 minute outage, and thus granted everyone as much api credit as they used the entire past week, which seems generous 0130 13
stellz @piss.beauty · 15/10/2026it tastes different it's not as good you're right. aluminum forever 060 6
stellz @piss.beauty · 15/10/2026no you should take a hostage and demand they provide free diet coke to employees 010 1
stellz @piss.beauty · 15/10/2026adam's taste in guests feels calibrated to make everyone slightly mad at him intermittentlyyoutube.comGRIMES Talks Elon Musk, AI, and the End of HumanityYouTube video by The Adam Friedland Show 2163 16
stellz @piss.beauty · 15/10/2026it's a bit of both, i wanted to try a more complex use case for this little model, see if i can get something that runs on a 16gb unified memory machine to actually be useful 110 1
stellz @piss.beauty · 15/10/2026hmmm looking at my posting lately i think i might be developing a bit of the madness 1100 10
stellz @piss.beauty · 14/10/2026it (experimentally) works with pdfs too it decomposes the pdf into text & images (including rendering diagrams to images) using poppler tools (pdftotext etc.) the qwen model is multimodal so you can input images & get those images considered in the summary 280 8
stellz @piss.beauty · 14/10/2026dedicated summarizer / compactor is the first useful thing i've found to do with this spare 16gb m1pro machine. with a bit of tuning qwen-3.5-9b quality on this can be decent if not amazing on longer contexts 1331 33
stellz @piss.beauty · 14/10/2026who had that tool that would take a bsky thread link and convert it into a text-only representation? am i imagining that someone posted this 480 8
stellz @piss.beauty · 14/10/2026the 32gb m6 chip would be such a budget sweet spot for smaller local stuff if they didn't make the memory bandwidth (170 GB/s) so much smaller than on pro/max series chips (last-gen m5 pro has 307 GB/s) especially since it seems there's not going to be m6 pro/max 1181 18
