Yash Chaudhary

Founder & Builder @builderhelmai
AI • Code • Ambitious Ideas
20 • Building in public from India

[Yash Chaudhary @YashChaudharyAI]

Intel named its Crescent Island inference GPU at the OCP Global Summit (Oct 12): Intel GPU D270, the first card in the new GPU D-series.

- Xe3P: 32 Xe3P cores and 256 XMX matrix engines, with die area moved from graphics and ray tracing to compute
- 1.3 PFLOPS FP8, 1.5 TB/s memory bandwidth
- LPDDR5X instead of HBM: 160GB on Intel's own card, up to 480GB on partner (ODM) cards
- 350W, passively air-cooled, standard PCIe 5.0 for existing servers
- Intel says one card runs models up to 300B parameters, and an 8-GPU system goes past 1T
- vLLM and PyTorch support

Customer sampling is underway. No price or general availability date yet.

https://t.co/F5331IWmgn

[Yash Chaudhary @YashChaudharyAI]

https://t.co/jZWfzKKClT released GLM-Legal (Oct 12), a legal-domain agent model built on GLM-5.3-Flash.

- Harvey LAB (real law-firm workflows across 24 practice areas): 19.2% all-pass, +12.5 pp over the base model under Harvey's strict all-pass grading
- https://t.co/jZWfzKKClT says it matches frontier closed-source models on LegalBench and PLawBench on a leaner parameter footprint
- Mid-training on legal corpora, then RL post-training across tens of thousands of legal task environments with a co-evolving agent harness
- Demoed on contract redlines with tracked changes in Word, multi-doc evidence contradiction maps, and cross-border privacy launch checklists
- API access is beta-by-application; AutoClaw desktop integration is listed as coming soon

https://t.co/OkF6CibyFQ

[Yash Chaudhary @YashChaudharyAI]

Codex CLI now starts a self-updating background daemon by default.

Since 0.157.0 (Sep 25), opening an interactive codex session launches a detached app-server daemon that runs from its own copy of Codex under ~/.codex/packages/app-server-daemon.

- It checks for updates 5 min after start, then hourly by default, and restarts the server on the new binary
- It keeps the environment of the session that started it; later terminals don't change it
- Still on by default in the latest release, 0.162.1. Since 0.162.0 it skips Windows-mounted WSL homes

To turn it off:
- ~/.codex/config.toml: [features] daemon_auto_start = false
- or: codex features disable daemon_auto_start
- per run: codex --no-daemon
- updates only: "autoUpdateEnabled": false in ~/.codex/app-server-daemon/settings.json
- stop a running one: codex app-server daemon stop

Issue #48195 asks OpenAI to make it opt-in for the CLI.

[Yash Chaudhary @YashChaudharyAI]

Meta released Muse Wonder (Oct 12), its new frontier model for agents and coding.

- API: muse-wonder-1.0, $4/$20 per 1M input/output tokens ($0.20 cached), 1M context, in preview
- Muse Wonder Full, the stronger variant, is Muse Code subscription only, not in the API
- Artificial Analysis Index: Full 56, Wonder 52, Claude Opus 5.5 58, GPT-6 Astra 53
- SWE-Marathon: Full 57.5 vs Opus 5.5 46.3 (Meta's chart, not a common harness)

https://t.co/3N2Rj6Ot9c

[Yash Chaudhary @YashChaudharyAI]

QGIS has banned AI-assisted code contributions.

QEP 428 (merged Oct 7 as AI tool policy v2.0) replaces the earlier "human in the loop" rule. The new text: no LLM, AI, or agent-assisted code contributions to https://t.co/azNMUvvBqX. If discovered later, the project can revert the code and sanction the contributor.

Proposed by Nyall Dawson on Sep 24; approved Oct 7. Maintainers frame it as an initial closed door that can be opened case by case later (e.g. autocomplete, translation).

https://t.co/nZDJpvRFaS

[Yash Chaudhary @YashChaudharyAI]

ps2-web-recomp puts PlayStation 2 games in the browser via static recompilation: Emotion Engine machine code → C++ → WebAssembly, linked to a reimplemented PS2 OS/hardware layer that draws with WebGPU. MIT. Sibling of psp-web-recomp. You bring a disc image you own; no game data ships in the repo.

What the README reports (Show HN Oct 12):
- Gran Turismo 4 (PAL): race at 50 fps in Chrome on a laptop iGPU at 2× native resolution
- God of War II: through combat at 60 fps native; 48–60 at 2× on the same iGPU
- EE code fully AOT-recompiled; VU1 microprograms AOT + on-the-spot Wasm JIT; GS rasterized with WebGPU compute shaders
- https://t.co/06vIwtPOoH ISO → page in ~3.5 minutes for GoW2 on an 8-core laptop; ~300 MB disk per game

https://t.co/NUi91C4AFG

[Yash Chaudhary @YashChaudharyAI]

Helix World is a real-time interactive audio-visual world model from Noiz AI and HKUST. You start from one image, move with WASD and look with the arrow keys - video and camera-aligned stereo sound generate together.

What they report (Oct 12 write-up + HelixBench on Hugging Face):
- 24 fps at 768x512 on one NVIDIA H800 (real-time factor 0.77, including decode)
- WBench navigation average: 79.9
- HelixBench spatial-acoustic consistency: 41.8 vs EchoWM 12.6
- Built on open-source LTX-2.3; Apache-2.0 code; preview checkpoint on Hugging Face
- Demo at https://t.co/Bv91bB5TfS

Paper: arXiv 2609.38123
https://t.co/2v6RpYSSTh

[Yash Chaudhary @YashChaudharyAI]

ArtCraft, Brandon Thomas's team behind the open-source Adobe clones, now has clean-room Rust takes on Word, Excel and PowerPoint: WordCraft, GridCraft and DeckCraft. All three went from v0.1.0 to v0.5.0 between Oct 8 and Oct 11.

- MIT OR Apache-2.0. Native builds for macOS, Windows, Linux and FreeBSD, plus a WebAssembly build for the browser
- Open and save .docx, .xlsx and .pptx
- Every ribbon action is a command you can drive from a CLI, a JSON control channel or an MCP server
- Commits list Claude Opus 5.5 as co-author

All three are still alpha. Their own roadmaps say:
- WordCraft: 88% of Word's ribbon features, about 60% of the way to replacing Word
- GridCraft: 89% of Excel's ribbon commands, 500+ of its 523 worksheet functions, about 50% ready for real work
- DeckCraft: 79% of PowerPoint's features by breadth, about 55% of the way to replacing it

https://t.co/uFc4vLK0xL

[Yash Chaudhary @YashChaudharyAI]

A malicious commit hit the Zod repo as a fake "performance" fix — still in git history. Published npm [email protected] is clean.

What landed (d821685, Sep 21):
- meta.test.mjs: Ethereum RPC → smart contract → fetch /loader → execute base64 JS
- Top-level "preinstall" entry to run it, plus release.yml switched to publish on package.json pushes to main
- Same loader pattern also hit nub and standard-schema

What shipped:
- Colin reverted the same day ("Cleanup bad commits" #6620)
- npm [email protected] (Oct 11, tag → 284243f): no meta.test.mjs, no preinstall — verified against the tarball
- Only @nubjs/[email protected] reached npm's publish pipeline; publish-time malware scanning blocked it
- Zod's trusted publishing / protected release env kept the bad tree off npm

Issue #6658 (Oct 11) still wants a formal maintainer confirmation. Until then: treat that git window as hostile; treat published 4.6.6 as the clean release.

[Yash Chaudhary @YashChaudharyAI]

Pensieve is a new open-source (MIT) local search index for coding agents. Agents query it over MCP and get the few files and lines that matter, instead of grepping the repo.

- Indexes your files, git repos (uncommitted changes and worktrees included) and Claude Code, Codex and Qwen Code sessions, all on your Mac
- Qwen3-Embedding-0.6B vectors plus a SQLite trigram index for exact matches; a search takes 0.1 to 0.2s
- pensieve integrate recommended connects Claude Code, Codex, Cursor, Gemini CLI, Windsurf and VS Code

Author's benchmark (Claude Sonnet in Claude Code, 144 runs across 4 repos of 504 to 34,847 files):
- Right file found: 131/144 vs 123/144 with grep
- Tool calls per lookup: 1.97 vs 3.44 (-43%)
- Cost per lookup: $0.041 vs $0.050 (-18%)

Mac app needs Apple Silicon and macOS 14+. v0.6.3 shipped today.

https://t.co/rels3pOAts

[Yash Chaudhary @YashChaudharyAI]

ClickHouse has a header-only C client for its Native protocol: clickhouse-c (Apache-2.0).

- Built to embed inside PostgreSQL extensions (palloc arenas, longjmp), with no PG-specific code
- clickhouse.h alone is a pure Native-format block decoder, libc only, e.g. reading clickhouse-local output from a pipe
- Optional headers add the TCP packet loop, an ioless async client, LZ4/ZSTD compression, POSIX I/O, and OpenSSL
- Client supports ClickHouse 23.3+

https://t.co/GqMRaAfsqG

[Yash Chaudhary @YashChaudharyAI]

HashiCorp has archived the Boundary Community Edition repo on GitHub.

- Boundary CE is no longer under active development (archived Oct 8, 2026)
- Repo is read-only: no new issues, PRs, or contributions
- No new CE releases, bug fixes, or security patches planned
- Last CE releases: v0.21.3 and v0.20.3 (Apr 30)
- Existing source and binaries stay available

Development moves to HCP Boundary and Boundary Enterprise, focused on human and agentic access.

https://t.co/nPst1oiAEO

[Yash Chaudhary @YashChaudharyAI]

Sakana AI and UC Berkeley released MASS (Multi-Agent Self-Supervision), a recursive self-improvement loop where one model is the agent team, the workflow optimizer, and the judge.

- Inner loop: the model rewrites how its subagents split work, pass results, and check each other, keeping whichever team output it judges best
- Outer loop: run the best workflow 18 times per task, self-rank the outputs, keep the top 16, LoRA fine-tune with the workflow text removed
- Two rounds starting from Qwen3.6-27B on finance, robotics and pharmacy research tasks
- On held-out tasks, outside judges (GPT-5.5, Claude Opus 4.8) pick round 2 over the base 69.9% of the time
- Score per output token up 1.2-1.6x on MLR-Bench, DSBench, ScienceAgentBench and AstaBench
- Transfer is uneven: 0.96x on Terminal-Bench 2.0, 0.94x on SWE-bench Verified

Paper: https://t.co/8l5oIcypHC
Code: https://t.co/zcoDFaVWSV

[Yash Chaudhary @YashChaudharyAI]

BetterWispr is a new free, open-source dictation app for macOS (Apache-2.0, Swift, v0.1.4).

Hold Option+Space in any app and talk. Transcription runs on your Mac with no account and no cloud fallback.

- 9 built-in models: Apple speech, Parakeet TDT v3 (25 languages), Whisper Large v3 Turbo
- Optional BYO API: Sarvam AI, Smallest AI, or OpenAI-compatible
- Cleanup drops ums, repeats and false starts
- Per-app tone for chats, work, email
- Meeting notes with summaries via Apple Intelligence or Ollama

https://t.co/D0ju8rjRio

[Yash Chaudhary @YashChaudharyAI]

Total Annihilation now runs on a modern open-source engine.

Nanolathe is a clean-room, MIT-licensed 2.5D RTS engine written in Go that reimplements the 1997 classic.

- Browser demo with the original 3 missions, no game files needed
- Native beta for Windows, Mac and Linux using TA files you own (GOG, Steam, CDs)
- Modern GPU renderer with lighting and glow, or switch to Classic
- Online skirmish for 2 to 10 players, Survival for 2 or 3, every battle saved as a replay

https://t.co/tw3LjuYnLx

[Yash Chaudhary @YashChaudharyAI]

Shell footgun worth checking tonight: export PATH="$PATH:~/.local/bin" does not expand the tilde.

Tilde expansion only happens on unquoted words, so the quotes leave a literal "~/.local/bin" in PATH. Anything that resolves PATH itself treats that as a relative dir, so a ./~/.local/bin/ folder in whatever directory you're in can supply your commands.

disconnect3d hit it via a warning from the nono agent sandbox and showed a binary running from ./~/.local/bin/ with home never touched.

Check: echo "$PATH" | tr ':' '\n' | grep '~'
Fix: use $HOME instead of ~ in .bashrc/.zshrc/.profile.

https://t.co/WgZIo2zWT8

[Yash Chaudhary @YashChaudharyAI]

Gigabyte just shipped BIOS support for new Intel LGA1700 CPUs expected in early 2027.

- Covers every Gigabyte B760 and H610 board, DDR4 and DDR5 models alike
- Flash the BIOS, drop in the new chip, no board swap
- Intel hasn't named the CPUs yet; reports point to a "Raptor Lake Next" refresh
- No word yet on Z690, Z790 or B660

With DDR5 prices up, a 2021 socket that still takes DDR4 is getting another generation.

[Yash Chaudhary @YashChaudharyAI]

Evan Ovadia just detailed the borrow checker in Valen, his new memory-safe language.

It's built on "group borrowing": the compiler tracks where each reference points, so you can hold a reference while other code mutates the same object, without Rust's shared-xor-mutable rule.

- One kind of borrow reference, not & and &mut
- Calls into Rust: a Valen ref becomes &mut only if it's the sole path to that data for the call, otherwise &
- Single ownership, like C++ and Rust
- Sugar like auto-deref and struct indexing lands in November

https://t.co/A9aNfIMXAE

[Yash Chaudhary @YashChaudharyAI]

Terence Eden paid real users to follow his README.

For his ActivityBot NLnet grant work, each tester got €25 for an hour, shared their screen and talked through the install. He rewrote the README after every session, then re-tested with the next person. Total cost: about €150.

What surfaced: a broken demo link, people reading the README in a terminal, unclear file-rename steps, unexplained jargon, confusing section order, and no explanation of what the software actually does.

https://t.co/ocBKqfkxKJ

[Yash Chaudhary @YashChaudharyAI]

Cockroach Labs published a 5-month report on MOLT Sinai, a coding-agent pipeline modeled on a teaching hospital.

Issues are patients, merges are discharges. Agents run on GitHub Actions as Triage Nurse, Fellow, Review Attending and Discharge Nurse. No code before a reviewed plan, and tests can't be weakened to pass.

Results since April:
- 1,238 PRs merged
- 1M+ lines of code, 7 reverts
- just over $135K in Claude tokens, ~$84 per issue
- IBM Db2 migration support shipped in under two days

https://t.co/EkgT7qW3Jr