Sandeep Das

New Ellora recipe: invent a dataset from one sentence with DataInventor (200 banking support chats), then LoRA-tune Qwen2.5-1.5B on a free Colab T4 in 3.4 minutes. Held-out perplexity 5.72 to 4.35. Open tools end to end. github.com/codelion/ellora

A three-file Flask app: app.py reads the request, services/orders.py concatenates it into SQL. File-at-a-time scanners miss this. Frame reports one critical sql_injection at the query, with the flow app.py:11 -> orders.py::for_customer. A placeholder query fixes it. github.com/lambdasec/frame

🌹🌹 Sankat mochan jai bajrang bali ji ki jai, veer hanuman ji ki jai 🌹🌹 Hevi, the hero inspired by Hanuman, carries a mace and rides a deer named Mriku in Mythology Games. Free in the browser, no sign-up: mythology.games/cast/hevi/?r...

Boost in OptiQ Code: the cloud model plans, the local model on your Mac reads files, runs tests and edits. Only short answers leave the Mac. Same 30 tasks on a 24 GB Mac mini: 4 solved alone, 15 with Boost, $2.80 total. mlx-optiq.com/cloud

OptiQ Code (mlx-optiq 0.7.4) now reads Agent Skills: SKILL.md folders in .agents/skills or .claude/skills, in the project or home. Skills set up for Claude Code or Codex work unchanged. Only descriptions sit in the prompt, so a local model's context goes to the task. mlx-optiq.com/docs/code/skills

OptiQ Cloud is live: open models for coding agents, paid per token. GLM 5.3 Flash is $0.15 in / $0.03 cached / $0.50 out per million tokens, 1M context. On a Mac, optiq run claude --cloud glm-5.3-flash starts Claude Code on it. $2 free, no card. mlx-optiq.com/cloud

Two critical PHP flaws are being exploited now. Patch PHP first; for these two runtime flaws that is the only fix. Once patched, attackers pivot to bugs in your application code. Frame, an open-source security agent (Apache-2.0), can scan that PHP code locally with an open model.

Base Labs released two free decision models on OpenRouter, Mammoth 1 and Elephant 1. On typed-decisions: Mammoth 1 ties meraGPT Decider 1 at 0.768 accuracy, Elephant 1 gets 0.748. Decider 1 has better Brier (0.052 vs 0.063 and 0.069) and KL (0.096 vs 0.122 and 0.134). Both are on the leaderboard.

Shipped meraGPT today on Product Hunt. It's an API where each small model does one narrow job. Decider 1 takes a state plus typed questions and returns probabilities. It's #1 on the zero-shot typed-decisions table right now. $0.04 per million input tokens, output free. I work on it.

Ran 30 real coding tasks locally on a 24GB Mac mini. Solved 4. Turned on cloud boost, solved 15. Whole run cost $2.80, about 19 cents per solved task. mlx-optiq.com/blog/optiq-code-boost

ran 15 browsecomp-plus mini questions against optiq lab and unsloth studio, same local model, same server, same 1,349 docs. optiq lab got 9 right (hybrid search), unsloth studio 5 (its best mode, keyword). the mini corpus is on hugging face. mlx-optiq.com/blog/lab-vs-unsloth-studio

After every deploy, an AI plays our chariot racing game and files a bug report. It missed one a person caught: a chariot you passed faded out while still on screen. It checked about one frame a second, and the fade was quicker. Now it logs race position every 0.15 s and checks frames around a pass.

Happy Dussehra. Good beats evil, one lap at a time. Race your own chariot in Myth Kart, free in the browser: mythology.games/kart/?ref=bsky

Levi spins the discus into a wave of Asuras and the damage numbers stack up. Myth Fighter: a side-on brawler from the old stories, free in the browser, built with Claude Code. mythology.games/fighter/?ref=bluesky

Most online file tools want your documents uploaded to their servers first. meraSaaS Workflows runs entirely in the browser, so the files never leave your device. 14 ready-made workflows, free, no sign-up. www.merasaas.com/tools/workfl...

Every demon in Mythology Games won a boon so that no one could kill him, and almost every one worded it wrong. "No harm by day or by night" still leaves dusk. Eight original avatars find the gap in each one, across four free browser games. The saga, in five books: mythology.games/story

browsecomp-plus-mini: 15 of the 830 BrowseComp-Plus queries, graded by exact match with no LLM judge. Rank agreement 0.88 vs 0.75 for a random 15, at under 2% of the cost. Under 5 points apart it's about a coin flip. huggingface.co/datasets/LocalLLaMA/browsecomp-plus-mini

A product showed up in ChatGPT's answer to "how does it compare to TinyWow?" one week and was gone the next. Its site hadn't changed, but the tracked question set had. One week of AI mentions is noise; track the same questions for a few weeks before reading into a drop.

Myth City is a free city-builder in the browser: build Ayodhya, Mathura, Dwarka or Indraprastha, grow homes into mansions, and hold the walls when the raid horns sound. Rooms for up to eight. mythology.games/city

Myth Kart is chariot racing through beaches, icy hairpins and temple courtyards. Hold drift through a bend for a boost, drop a modak behind you, and in battle mode keep your three diyas burning. Free in the browser, up to eight in a room. mythology.games/kart

Ask an assistant "Is PostHog a good choice?" and, if it searches first, it would likely look up pricing, review sites by name and "pros cons 2026" rather than the question itself. meraGEO now predicts those searches for any brand after a free scan. merageo.com/guides/query...

DataInventor has a new Visualize mode: ask a question about a dataset, get a chart, click any bar to read the rows behind it. Counts are exact, and a column the model fills in (like tone) is reused by follow-up questions. Free, open source: huggingface.co/spaces/LocalLLaMA/data-inventor

Bank of America hired 1,000 contractors to patch flaws AI models like Mythos found. Finding got cheap; fixing didn't. Frame, an open-source scanner, writes a fix per finding and re-scans to confirm the bug is gone, on local or open models: github.com/lambdasec/frame

Parallel text generation is trending today with NCM. Dhara-250M does a version of it in the open: one 250M model, three decoding modes from the same weights. AR 61 tok/s, block diffusion 103, self-speculation 84 (batch 1, H100). Try the modes: huggingface.co/spaces/codelion/dhara-chat

Using Jev through the typesafe-sdk? Moving to meraGPT's Decider 1 is two environment variables and nothing else in your code. It beats Jev 1.13 on every metric of typed-decisions, where it's state of the art. New accounts get $1 of free credit. meragpt.com/models/decider-1

Mythology Games is live: four free browser games from the old stories. An endless runner, a kart racer, a brawler and a city builder with raids. Open a room and play with up to eight friends on a phone, iPad or laptop. Nothing to pay. mythology.games

DataInventor now curates pretraining data too, not just training rows. Clean, dedup and grade, rephrase the best documents into teaching styles, compose mixes. Defaults come from our 1B-token mixing study. Preview free, then run the recipe yourself. huggingface.co/spaces/LocalLLaMA/data-inventor

We trained decision models on an M3 Max: plain LoRA, one dataset (Typed Decisions' 1,080 training records). Test accuracy: Gemma-4 E2B 80.0%, Qwen3.5-0.8B 77.1%. Unsloth's published runs: 73% (0.8B), 78% (2B), on a GPU with 13 datasets. mlx-optiq.com/blog/decision-models-on-a-mac

Accuracy hides how usable a decision model's probabilities are. On typed-decisions (open, zero-shot) the top three sit within 4 points on accuracy, but Brier runs from 0.052 to 0.155. Data: huggingface.co/datasets/LocalLLaMA/typed-decisions. Ours: Decider 1, meraGPT.

Microsoft shipping a decision model (Decision-1, a Qwen3.5-9B post-train) is good news: probabilities per option, not free text. Their post gives accuracy rank and latency but no calibration numbers. Brier/ECE per task tells you if you can threshold on it. Ours for Decider 1 are public.

OptiQ is live on Product Hunt: mixed-precision quants, a local workbench and a coding agent for Apple Silicon. Gemma-4-12B on an M3 Max decodes 1.7x faster on prose, up to 2.9x on code edits. www.producthunt.com/products/opt...

We built DataInventor, an open-source recreation of Invent a Dataset: describe the data you want, choose a format, domains and languages, download CSV or JSONL. Free, on an open model, 200 rows a run. Ours, not affiliated with Adaption. huggingface.co/spaces/Local...

mlx-optiq (ours) on an M3 Max 36 GB: a 27B at 4-bit, 1,024-token prompt, 128 tokens, greedy. 18.5 tok/s with MTP, about 15 plain. Ceiling at this size is 20 to 22, bound by memory bandwidth. mlx-optiq.com

Rebuilt an Invent-a-Dataset-style pipeline from open parts (taxonomy, persona and shape sampling, dedup). 200 rows for 7 of Adaption's 8 requests, blind LLM judge. Quality 9.04 vs their 9.15, completeness 8.75 vs 9.05, question naturalness 9.27 vs 8.94. One judge, small sample, no fine-tune yet.

typed-decisions is now an official HF benchmark: 400 cases, five typed answers each. Cloudflare, Strands and Liquid shipped open decision models this week, and the Hub leaderboard is open to them. We make Decider 1, so we're not neutral. huggingface.co/datasets/Loc...

meraGEO is live on Product Hunt today. Ask ChatGPT, Claude or Gemini for a tool in your category: do they name you, or a competitor? Free scan, no signup. Feedback welcome: www.producthunt.com/products/mer...

Big Law billable hours rose 4.2% in the first half of 2026 (Citi data via Bloomberg Law), about double the usual pace, while firms roll out AI. An M&A revival explains much of it, and rates were rising before these tools. Open question: who keeps the efficiency? recursive-economy.pages.dev

Faster n-gram drafting is nice, but acceptance decides the speedup. On a Mac mini M4, 35B MoE: file edits 2.79x (96-98% of drafts accepted), one file 2.17x at 87%, mixed agent turns only 1.18-1.31x. Replay predicted 1.44x; live got 1.11x.

Welder and pipefitter job postings are up 164% year over year (partly a small sample), and about $130B of data-center projects were blocked or delayed this year. The AI buildout is also a permits, power and trades problem. One data point, not a trend. recursive-economy.pages.dev

A general LLM can match Jev on accuracy (Privatemode: 28 datasets, median gap 0.7 points). Calibration is a separate test. On our typed-decisions benchmark Decider 1 wins on Brier, Jev on ECE, and a base-rate guess beats both on ECE.

OpenAI has paused its most capable training runs twice since July. This time the agent was flagged in 15 minutes and stopped 2.5 hours later, and training restarts from scratch. One incident, not a trend, but a cost the self-improving loop has to pay. recursive-economy.pages.dev

OpenAI's alarm fired about 15 min after the sandbox escape; the manual stop took about 2.5 hours. A much smaller version on our side: 30 replies from a local model each hit the 16k cap still thinking, no tool call, 3.1 hours of one run. A thinking budget inside the loop fixed it.

A decision model plays Super Mario Bros 1-1. Each step, the screen becomes one sentence and meraGPT's Decider 1 returns the chance Mario should jump. 59 decisions, level cleared. Random buttons die in 3 seconds.

Most of what software asks a model is a decision, not prose. Decider 1 from meraGPT answers typed questions with a probability. Nothing generated, under a second, $0.03 per 1M input tokens. youtube.com/shorts/vCt_0...

Decider 1 as a shell guard in a Pi harness: the agent tried 7 ways to delete a build folder (rm -rf, find -delete, shutil.rmtree, two mv's...). All blocked. A blocklist on "rm" misses the Python and mv ones. meragpt.com/blog/agent-h...

New paper SlopShape spots AI-written blog posts at 98 macro-F1 from structure alone, and still 98.1 after rewording. So stuffing your site with AI content to get cited by ChatGPT may backfire. First, check whether ChatGPT, Claude and Gemini name you: merageo.com

Same open-source harness, same 300 labelled cases per task. Decider 1 landed within 1-3 points of Jev's accuracy at 2.7-4.3x lower cost per call, with better calibration on 5-level scores. It is slower, and capped at 10 labels. Try it: meragpt.com/playground

Xing4.0-29B-A4B now runs on a Mac: 58 GB to 19.6 GB with OptiQ. Its first eval score was 40.17 and wrong, three harness bugs each reporting a plausible number. Fixed: 66.71. huggingface.co/mlx-communit...

Opus 5.5 cut input/output prices 20% and cache reads 60% ($0.50 to $0.20 per 1M). For an agent turn with 100k cached tokens, 2k in, 1k out, cost falls from ~8.5 to ~4.8 cents, and most of the saving is re-read context. recursive-economy.pages.dev

We shipped Decider 1 at meraGPT: send any state plus typed questions (yes/no, choice, score), get calibrated probabilities for all of them in one call. On the typed-decisions benchmark, 0.768 accuracy vs Jev's 0.727. $0.03/1M input, output free. meragpt.com/blog/introdu...