Newsletter
![The image displays a stylized world map with glowing network lines connecting various points, overlaid with numerous small rectangular boxes containing images of people and text labels. Prominently in the center, large blue text reads "AITINKERERS" and "POST-TRAINING", suggesting an informational or promotional graphic.. Text: AITINKERERS
"POST-TRAINING"
PALLADIUM
ELENIOM
BANDICAM
GARDENWARE
MOI AFARLE
ENIGMA
NO FARM IN RAIN CINE PLENIUR
Mianny Hotel
FREELANCE
ALPHONSO
S. Africa
THE FUGA
AL PRUUCE MAGNA COREM ABDUILE BERLLINE ME CEULANTEPIC AN
[numerous other small, unreadable labels and names]](https://images.aitinkerers.org/cdn-cgi/image/width=700,fit=cover,format=auto,quality=68,height=475/blog_images/post-training-screen-png-7AO1.png)
Explore Top AI Demos #44, highlighting practical breakthroughs in AI avatars, voice coding agents, and physical robot perception. Expect a mix of hands-on systems and real-world sensing work—plus the community energy to see how these capabilities come together in demo form.
Top AI Demos #43 brings practical updates you can try: adaptive math that responds to problems, agent tool security-focused work, and local search techniques for faster, more relevant results. Thanks to everyone building these systems—great energy for hands-on experimentation.
Explore this week’s top AI demos: compiling 1T-parameter LLMs down to 100MB, adding video memory for more capable agents, and setting up local voice dashboards. These practical updates focus on pushing model size and agent memory forward—thanks for sharing and building with the community.
Explore Top AI Demos #41: a Live Media Studio for real-time AI workflows, robotic pest control showcasing autonomous action, and an AI product launch highlighting how demos translate into practical releases. Thanks for joining the community—there’s plenty here to spark hands-on experimentation.
In Top AI Demos #40, explore zero-cost companies, declarative agents, and LLM world sims—three practical threads worth tracking. If you’re building or researching, this mix highlights new ways to structure AI work and simulate environments, with enough novelty to spark experiments.
Check out the newest AI demos: AI SEO engines, agent memory attacks, and hardware-focused AI experiments. If you care about practical SEO automation, agent security, or how AI runs on real hardware, these updates are worth your time—thanks for keeping the community engaged.
After Anthropic suspended Fable 5 (June 12–July 1, 2026), we built an antifragile billing agent that keeps working when models/providers fail by routing through OpenRouter presets and evaluated fallbacks. Using TypeScript code for money/policy logic and a 20-case synthetic eval (8 holdout), Grok 4.5 matched all decomposed cases (20/20) at 63% lower cost than Fable 5 ($0.0555 vs $0.1492) and Grok became primary while Fable stayed fallback. We store model order in an OpenRouter preset (no deploy needed) and use Router metadata/Broadcast to trace which model answered.
In Top AI Demos #38, you’ll learn how to build inspection VLMs, run local-first agents, and use agent harnesses to test reliably—plus practical demo takeaways to help you ship faster with confidence. Thanks to the community for sharing these hands-on approaches.
Giving an avatar “a face” bottlenecks on interaction timing, not rendering: interruptions, turn-taking/ending cues (e.g., Turkish needed eager end-of-turn disabled), visible thinking time, and true audio-driven lip sync. With Anam bring-your-own-LLM, keep the system prompt short (1–2 spoken sentences) and stream Fable 5 tokens into the talk stream; Anam targets ~180ms response.
Top AI Demos #37 brings practical agent-building updates: Pi-Hydra agents, ready-to-use agent guides, and AI hardware design. If you’re working on agent workflows or pairing models with hardware, this roundup highlights the most actionable developments from the demo set—thanks for building with us!
Explore Top AI Demos #36 for practical local AI servers, forkable sandbox setups, and agent memory ideas. If you’re building or experimenting, these demos share hands-on paths to run AI locally, extend sandboxes, and improve agent context with memory—great food for community tinkering.
Top AI Demos #35 highlights fair agent evaluations, research agent sandboxes, and “MRI for LLMs”—practical updates aimed at making agent testing more equitable, giving researchers safer sandbox environments, and improving how LLMs are evaluated.
This week’s Top AI Demos #34 highlights how on-device LLMs are evolving, shares fresh Agent Lab report takeaways, and explores practical store automation. Thanks for tuning in—there’s a lot of usable direction here for builders looking to ship AI that runs closer to the device.
Check out Top AI Demos #33: LeRobot’s Connect Four robot, Jaxpot’s 100M steps/sec RL training, and Claude Code’s 3D game capture. If you like hands-on systems and measurable performance, these updates are a great place to dig in.
If adding instructions makes your agent worse, it’s often a harness problem: we watched Reef/AlphaCumen collapse from a Vals v1 lead to zero on Vals v2 because prompts exceeded 2,500 lines and 50+ overlapping tools, causing token-budget timeouts. Reef fixes this with git-native, skills-first ReAct: skills bound by `SKILL.md` + `@skill_fn`, lazy-loaded skill indexes, planner specialist routing, and runtime constraints like `inject|clamp` as-of cutoffs. Result on finance evals (n=239): 82.6% pass rate with full structure vs 44.87% reference and 49.8% retrieval-only, at ~10× lower cost (620k tokens/query; $0.13 vs $1.35 at Opus 4.7).
Top AI Demos #32 brings actionable experiments: Cloud Waste Agents for real-world problem solving, a practical MVP validation workflow, and AI skill orchestration to coordinate tasks effectively—great insights for teams building and iterating quickly.
This week’s AI demos focus on practical on-device AI, agent-based code search, and phone coding: discover how models run locally, how agents help locate relevant code faster, and how coding workflows translate to your phone—thanks to the community for sharing these tools and ideas.
Learn to run open-source LLMs locally on Mac Apple Silicon with Apple’s MLX-LM: install `pip install mlx-lm`, then `load()` a Hugging Face model and call `generate()` (optionally `stream_generate()`); examples show ~35.5 tok/s on an Air (smollm3-3b-mlx, ~2.6 GB peak) and ~79 tok/s on M4 Max (gpt-oss-120b-mlx, ~63 GB peak).
This roundup highlights three AI directions: clinical co-pilots for healthcare workflows, “pull agents” that proactively retrieve needed information, and life planners that translate goals into actionable plans. Thanks to everyone building these tools—share your real-world wins and what’s working.
In this week’s top AI demos (#29), explore three practical highlights: agent memory for smarter ongoing context, image relighting for dynamic lighting changes, and software testing workflows that help validate results faster. Thanks to everyone sharing these demos—great ideas to try and build on.
Top AI Demos #28 brings practical projects to the community: real-time audio classification for instant listening insights, federated learning privacy that keeps training data protected, and interactive presentations for hands-on exploration. Thanks to everyone building and sharing—there’s plenty here to try, learn from, and extend.
Top AI Demos #27 shares practical builds: self-improving agents that iterate on their own, plus Musebot and Thrum demos to explore what these agent systems can do. Big thanks to everyone in the community driving hands-on experiments and sharing learnings.
Homecrew is an open-source package manager for agent skills: a skill is a directory with a `SKILL.md`, and a tap is a git repo or local directory of skills. Install with `crew tap add …` + `crew install`, then `crew update` keeps skills synced across all detected agents (e.g., `~/.agents/skills/` and agent-specific dirs), including dependency installs and automatic tap-level updates via `crew autoupdate enable`.
In this edition of Top AI Demos: PostHog MCP delivers full product visibility, self-evolving agents, and pooled compute so teams can observe and iterate faster. Parthas brings multi-agent Claude Code for collaborative code execution—great for practitioners who want concrete workflows, not theory.
Explore three practical AI updates: voice agents, MLX trading cards, and cyber threat intelligence. Dive into how voice agents work, what’s new with MLX trading cards, and the latest threat intel insights—then share your take with the community.
In Top AI Demos #24, explore deterministic agent pipelines and how “agent memory” changes behavior, plus a careful look at code leaks. If you build or test agents, these demos are practical prompts to tighten reliability, observe memory effects, and protect code. Thanks for bringing these insights to the community.
⚡ Explore 24/7 Dev Agents paired with MCP Apps—always-on agents that keep working and MCP apps that integrate smoothly—so your workflow stays active around the clock. Thanks for building with us and sharing what you’re testing!
⚡ Autonomous software factories pair well with agentic security to help teams ship faster with safer builds. The key idea: use agents to automate parts of the software pipeline while applying security checks along the way, so reliability and protection stay built-in as work progresses.
Tinkering at Scale focused on agentic orchestration, knowledge graphs, and creative AI, highlighting practical ways to coordinate complex workflows and connect structured knowledge for better outcomes. The session shared concrete wins from creative applications, and the community energy made it easy to translate ideas into experiments.
Agent version control plus pipeline-parallel LLMs help teams iterate safely while scaling throughput. By tracking changes across agent versions and running models in parallel stages, you reduce risk and speed up experimentation—so workflows stay consistent as you push new ideas into production.
WebMCP is coming to Chrome, alongside a 3D “Codebase Cities” view to explore code visually. The stack pairs with self-improving coding agents that iterate over your codebase—helping developers move faster with clearer context and tighter feedback loops.
AI Tinkerers #18 spotlights MemoryLane’s screen context and on-device voice models—sharing how to use screen context effectively and keep voice processing on your device. Join the community discussions and learn practical approaches for building and experimenting with on-device voice experiences.
After three months deploying OpenClaw beyond demos, the workflows that stick follow one shape: live inside tools you already use, carry context forward, and save time without babysitting. I built four systems—Robin’s nightly arXiv research (silence on weak results), a two-layer knowledge system (Obsidian facts + Honcho decision reasoning queried together), SnapshotClaw versioning for ~/.openclaw (configs/cron/skills/memory/plugins) to recover after 2AM breakage, and a Reachy Mini “brain” via Botwick that made the robot useful the next day (mic→VAD→STT→Claude→TTS).
In AI Tinkerers #17, we build multi-agent Openclaw comics and tackle security considerations while adding an AI fact-checker. Thanks to the community, we’ll share practical notes on prompting and safety—plus how the fact-checker helps keep claims grounded.
AI Tinkerers #16 dives into vector database evolution and persistent agent memory. The session highlights practical shifts in how embeddings are stored and retrieved, plus strategies for keeping agent context across sessions—so builders can design more reliable, stateful AI systems.
Community Spotlights • Issue #15.
Community Spotlights Issue #14 highlights reliability, offline safety, and product-ready AI builds.
Community Spotlights • Issue #13.
Community Spotlights • Issue #12.
Leverage air-gapped OCR with DSPy Vision to build better agents.
⚡ AI Gatekeepers + Legislative Monitoring with Gen AI highlights how Gen AI enables gatekeeping and legislative monitoring.
⚡ RapidFire combines Hyperparallel Tuning with AI Code Tracking.
This post covers Claude Draws, the Second Brain API, and more.
This post highlights On-Device Apple VLM and Windows Speech 🚀.
Image downscaling attacks and an open-source benchmark for app gen offer practical, reproducible research for the community. This update presents hands-on guidance, shared tooling, and benchmarks to help researchers and builders strengthen generation apps together.
⚡ Prompt Hijack + Deterministic Co‑Pilots — See How.
![The image is a screenshot of a social media post, likely from X (formerly Twitter), featuring a user's profile, a long text post discussing 'evals' in the context of AI and coding companies, and an embedded preview of a video or web page with a transcript.. Text: swyx
@swyx
Subscribe
Claude Code: no evals
[well known code agent company]: no evals
[well known code agent company 2]: kinda halfassed evals
[leading vibe coding company]: no evals
[ceo of company selling you evals]: mmmmm yess all my top customers do evals, you should do evals
[vc's in love with ceo of evals company]: mmmmm yess all my top founders do evals, must do evals
(NOTE: i -do- also think that evals are impt, but the eval pilled ai engineers have also noticed that it is not a strict requirement for success and, at least for 0-to-1 stage, may even be anticorrelated, think thru why)
(4) The future of agentic co...
New Chrome available
Transcript of The future of a...
youtubetotranscript.com
Premium
Search
+ Create
1:00
evals
- Yeah, so the best eval in some sense is
the one that most looks like real life.
And in that case, just using it gives you the
best result.
- We tried really hard, when building Claude
Code, to build a product evals.
- Yeah](https://images.aitinkerers.org/cdn-cgi/image/width=700,fit=scale-down,format=auto,quality=68/blog_images/swyx-png-6O-h.png)
Voice agents are hard because most use cases are long, multi-turn conversations and people don’t talk like they write, so LLM performance degrades over turns. Use context engineering + state machines (system prompt, conversation summary, limited tool list, next states). For confidence: do earnest manual testing, then lightweight/repeatable evals. I benchmarked a 30-turn voice scenario with ~75k-token context: every model had errors.
This week’s top demos feature Swift-native agents, micro:bit MPC, and Infinite Brainrot—an inviting snapshot for the community to explore these projects.
Local AI brings browser agents and ultra-low-latency meeting assist, delivering fast, client-side AI support directly in meetings.
![A screenshot of a dark-themed integrated development environment (IDE) displaying Rust programming code, with line numbers and syntax highlighting. On the right, a communication thread is visible, discussing the code's indentation levels and an embedded image of code.. Text: alex
1 month ago in # random - image.png
executor.rs
task_attempts.rs
claude.rs
amp.rs M
db.sqlite
backend > ~/bloop/vibe-kanban/backend/src/executor.rs tor for AmpExecutor > normalize_logs > [@] proce
26 impl Executor for AmpExecutor {
91 fn normalize_logs(
130 let processed = if let Some(msg_type) = json.get("type").and_then(|t|
entries.push(
timestamp: None,
entry_type:
NormalizedE
tool_nam
.to_
action_t
}
Thread
alex Jul 10th at 3:34 PM
amp.rs has 20 levels of indentation
image.png
1 reply
Louis Knight-Webb Jul 10th at 3:49 PM
Probably the most AI generated part of the codebase
Reply...
Also send to # random](https://images.aitinkerers.org/cdn-cgi/image/width=700,fit=scale-down,format=auto,quality=68/blog_images/image1-png-bnme.png)
Vibe Kanban showed how to ship a real CLI-built app: we capped PRs to <400 LOC and <10 files with a dedicated planning step (affected files, schema changes, test plan), used ts-rs to enforce shared Rust↔TS types via CI regeneration, ran agents in fresh per-session SQLite dev environments, and enforced separation of concerns with Rust crate boundaries (API can’t import SQL/git).
Post-Training Fellows
Have a bold AI idea, credible proof, and a hard-won lesson serious builders can use? Pitch a practical Deep Dive for 130K+ leading AI practitioners globally.
