{AI Marketing Playbook}

← All examples

AutoKernel (RightNow AI)

Research automationBreakout · Founder X post

Overview

Open-source Python framework that runs an autonomous coding-agent loop (edit one file, benchmark, keep or revert, repeat) to optimize GPU kernels for arbitrary PyTorch models, producing Triton or CUDA C++ kernels. Current scale: 1,407 GitHub stars / 142 forks / 12 open issues as of 2026-06-15 (per GitHub API). It is one of several viral OSS drops by RightNow AI (Jordan-based, Y Combinator F26), whose org also has openfang at ~17.8K stars and picolm at ~1.6K stars, so AutoKernel is a mid-tier hit in their portfolio, not their biggest.

First public appearance

Earliest verifiable public appearance is the GitHub repo itself, created 2026-03-11 00:12 UTC. The first distribution event was NOT Hacker News but founder Jaber Jaber's X launch tweet as @Akashi203, posted ~2026-03-11 00:55 UTC (about 43 minutes after repo creation): "i open-sourced autokernel -- autoresearch for GPU kernels. you give it any pytorch model. it profiles the model, finds the b[ottlenecks]..." This tweet was the breakout: 1,774 likes, 173 retweets, 38 quotes, 1,506 bookmarks (metrics captured via the @rightnowai_co retweet of it, RT, Mar 11 2026). The Hacker News submission AutoKernel: Autoresearch for GPU Kernels followed ~7 hours later (2026-03-11 07:42 UTC) by user "frozenseven" (not the founders), linking straight to the GitHub repo; it reached 47 points and 10 comments. So the founder's own X account, not HN, carried the launch reach (see revised Traction inflection). The KSP they led with (taken verbatim from the README headline, recovered via Wayback snapshot 2026-03-12): "Autoresearch for GPU kernels. Give it any PyTorch model, go to sleep, wake up to optimized Triton or CUDA C++ kernels." The framing immediately anchored to a borrowed-credibility hook: "Inspired by @karpathy/autoresearch -- which demonstrated autonomous AI agents for LLM training research. AutoKernel applies the same philosophy to GPU kernel optimization." Format: GitHub README as the canonical landing page (no separate microsite; the "homepage" field points to rightnowai.co/forge, their commercial product). The README at launch already included a Discord badge, a progress.png chart, concrete benchmark claims, a one-command quick-start (uv), and a KernelBench-integration section. By the first Wayback capture (2026-03-12 04:22 UTC, ~28 hours after creation) the repo was at 377 stars / 26 forks / 37 commits.

Launch sequence

  • 2026-03-11 00:12 UTC: Repo RightNow-AI/autokernel created on GitHub (MIT license). github.com/RightNow-AI/autokernel. Visible response: rapid star accrual within hours.
  • 2026-03-11 ~00:55 UTC: Founder Jaber Jaber (@Akashi203) posts the X launch tweet, "i open-sourced autokernel -- autoresearch for GPU kernels," ~43 min after repo creation. It became the single biggest distribution event: 1,774 likes / 173 RT / 38 quotes / 1,506 bookmarks (via the @rightnowai_co RT, Mar 11 2026). The company account @rightnowai_co amplified it by retweet rather than original post (its standing pattern, see Channel map).
  • 2026-03-11 07:42 UTC: Hacker News submission by "frozenseven" linking the repo. HN item 47332688. Response: 47 points, 10 comments; technical debate (comparisons to AlphaEvolve, TVM Ansor, llama.cpp integration ideas) plus benchmark skepticism (commenters "ademeure" and "aviinuo" challenged the matmul-vs-cuBLAS and CUTLASS numbers). The founders did NOT participate in the thread. HN was a secondary, smaller channel relative to the founder's X tweet.
  • 2026-03-12 03:22 UTC: Founder (@Akashi203) posts an update tweet ("autokernel just got a big update: native cuda c++ backend with 9 starter kernels..."), 67 likes via the @rightnowai_co RT (Mar 12 2026). Sustains the launch-week attention.
  • 2026-03-12 04:22 UTC: First Wayback capture: 377 stars / 26 forks. Wayback snapshot.
  • 2026-03-13 13:04 UTC: Founder posts "autokernel just took #1 on vectorsum_v2, 44.086us on B200" (60 likes via @rightnowai_co RT, Mar 13 2026), a competitive-leaderboard proof point feeding continued buzz.
  • 2026-03-12 03:16 UTC: Cross-posted into the NVIDIA Developer Forums (DGX Spark / GB10 board) by community user "joshua.dale.warner," framing it as "an adaptation of Karpathy's Autoresearch... worth experimenting with running overnight." NVIDIA forum thread. Low engagement (2 likes, 1 reply): niche but shows organic spread into hardware-dev communities.
  • 2026-03-19: Last code push to the repo (pushed_at 2026-03-19 03:13 UTC per GitHub API). The project effectively went into coast mode ~8 days after launch; the team's active commercial focus shifted to Forge / RunInfra.
  • 2026-03-22: arXiv paper v1 submitted: "AutoKernel: Autonomous GPU Kernel Optimization via Iterative Agent-Driven Search", authors Jaber Jaber and Osama Jaber (cs.LG, cs.PF). This formalized the launch and seeded a second media wave. Also mirrored on HuggingFace Papers and ResearchGate.
  • 2026-04-05 13:19 UTC: Founder (@Akashi203) posts the arXiv-paper announcement on X ("we published autokernel on arxiv, inspired by @karpathy 's autoresearch, we applied the same keep/revert agent loop to GPU ke[rnels]"), the second-biggest X moment for the project: 659 likes / 80 RT / 10 quotes / 636 bookmarks (via @rightnowai_co RT, Apr 5 2026). The explicit @karpathy tag again does the borrowed-authority lift. This X post, not the secondary tech-press writeups, was the largest amplifier of the paper.
  • 2026-04-06: Secondary media wave: MarkTechPost feature, Awesome Agents (updated 04-08), neurotechnus, and a Medium explainer by Aadish Agrawal. Several cite "roughly 1,000 GitHub stars within hours" and quote HN commenters, confirming the HN thread as the originating event feeding the media cycle.
  • 2026-04-12 14:07 UTC: Wayback capture: 1.2k stars / 114 forks; README now carries a v1.0.0 to v1.3.0 changelog (AMD ROCm support added, HF Kernels export, CUDA C++ backend, KernelBench) and an enterprise CTA ("Built by RightNow AI. For enterprise GPU optimization, check out RightNow Enterprise"). Wayback snapshot.
  • 2026-05-09: Founder Osama Jaber submitted the arXiv paper to HN (item 48074520): 4 points, 0 comments. Flopped.
  • 2026-05-13: Founder posted an explicit "Show HN: AutoKernel" (item 48120897): 2 points, 0 comments. Flopped. The founders' own promotional attempts dramatically underperformed the community-originated March 11 post.
  • 2026-06-15: Current: 1,407 stars / 142 forks. Growth is heavily front-loaded (see inflection).

Channels & accounts

GitHub repo: github.com/RightNow-AI/autokernel: 1,407 stars, 142 forks, 12 open issues, MIT, Python 100% (2026-06-15).
GitHub org: github.com/RightNow-AI: bio "GPU AI Code Editor," location Jordan, contact jaber@rightnowai.co, 527 followers, 18 repos. Notable sibling repos and stars: openfang ~17,826, picolm ~1,649, qwen3.5-triton ~115, rightnow-cli ~110, RightNow-GPU-Database ~87, AutoMegaKernel ~38, TIDE ~31, StreamIndex ~20. (The serial-OSS-launch pattern is the org's growth engine.)
X / company: @rightnowai_co: "RightNow (YC F26)," bio "Enabling Model-Hardware Co-Design at Scale | try: @runinfrai," verified, 1,002 followers, following 10, 62 posts, joined March 2025 (predates AutoKernel by ~1 year; created for the editor product). KEY PATTERN: the company account rarely originates; it overwhelmingly retweets founder @Akashi203's personal posts. Its own original product tweets are low-engagement (RunInfra launch Apr 7: 6 likes; TIDE research Apr 19: 9 likes; Forge upgrade Jan 23: 8 likes), whereas its RTs of @Akashi203 carry the real reach (the autokernel launch RT: 1,774 likes; the openfang OS RT: 4,322 likes).
X / founder (de facto megaphone): Jaber Jaber posts as @Akashi203 - now CONFIRMED live via the @rightnowai_co retweet timeline (Apify, Jun 2026). This personal account, not the company account, is the project's primary distribution engine: it authored the Mar 11 autokernel launch tweet (1,774 likes), the Apr 5 arXiv tweet (659 likes), and the org's biggest hit overall, the Feb 26 2026 openfang "operating system for ai agents" tweet (4,322 likes / 489 RT / 5,369 bookmarks, via RT). Co-founder Osama Jaber posts as @OsamaKakashi (confirmed via his May 31 2026 YC-announcement tweet RT'd by the company); he is the commercial co-founder and the one who posted the later (failed) standalone HN submissions.
X / sister product: @runinfrai (RunInfra, YC F26): their managed-GPU-inference commercial product, the current go-to-market focus.
Website: rightnowai.co (research lab framing); product surfaces Forge (rightnowai.co/forge, the AutoKernel "homepage" link), RightNow Editor, RunInfra (runinfra.ai); docs at docs.rightnowai.co; blog at rightnowai.co/blog. YC profile: ycombinator.com/companies/rightnow (batch F26, 2 founders, primary partner Nicolas Dessaigne).
Discord: discord.gg/UfEyc72t (linked from the README at launch via a "Discord - Join us" badge; current member count not visible).
HuggingFace: paper page huggingface.co/papers/2603.21331; AutoKernel also ships an export-to-HF-Hub feature (export_hf.py) but no prominent owned HF org kernel collection was confirmed.
YouTube: a video titled "RightNow AI Unveils AutoKernel: Transforming GPU Optimization for PyTorch" exists (watch?v=09seLc2LtsU); uploader/owned-vs-earned could not be confirmed (see Gaps).
Product Hunt: RightNow AI has a Product Hunt presence for the editor (not AutoKernel-specific).

Amplification & KOLs

Primary amplifier was founder @Akashi203's own X account (owned, not earned): the Mar 11 launch tweet (1,774 likes / 1,506 bookmarks) and Apr 5 arXiv tweet (659 likes / 636 bookmarks) far outweighed every other channel. Correction to the earlier read: Hacker News was a *secondary* amplifier, not the primary one. The Mar 11 HN submission by "frozenseven" (47 points) helped, but it landed ~7 hours after, and at a fraction of the reach of, the founder's X tweet.
Hacker News (community/earned): the March 11 submission by "frozenseven" carried a secondary slice of the launch. No named high-profile KOL endorsement was found; this was founder-account-driven plus crowd amplification, not third-party-influencer-driven.
Borrowed-authority hook (not amplification but central to positioning): the project explicitly framed itself as a domain transfer of Andrej Karpathy's "autoresearch." Karpathy did not endorse it, but the association did the lift in the README, HN title, NVIDIA-forum cross-post, and nearly every secondary article.
Earned tech-press / aggregators (all third-party, ~early April): MarkTechPost, Awesome Agents (byline Sophie Zhang), neurotechnus, aidevsetup, earezki.com (Dev|Journal), and a Medium explainer by Aadish Agrawal. Several recycled the "~1,000 stars within hours" line and the HN benchmark debate.
NVIDIA Developer Forums cross-post (community member, organic, low-engagement).
No evidence of paid promotion, sponsored placements, or paid KOL deals. Posture reads as fully organic/earned.

Traction inflection

REVISED after recovering the launch-window X data. The inflection was the founder's own X launch tweet (@Akashi203, 2026-03-11 ~00:55 UTC, "i open-sourced autokernel -- autoresearch for GPU kernels"), which hit 1,774 likes / 173 RT / 1,506 bookmarks and drove the repo from 0 to 377 stars within ~28 hours and to roughly 1,000 stars "within hours" per multiple earned writeups. The Hacker News submission (HN item 47332688, 47 points) was a *secondary* contributor, landing ~7 hours later and at far lower reach. EVIDENCE: (1) the founder X tweet posted 00:55 UTC, ~43 min after repo creation (00:12 UTC) and ~7 h BEFORE the HN post (07:42 UTC), with 1,774 likes / 1,506 bookmarks (RT); (2) Wayback snapshot 2026-03-12 04:22 UTC shows 377 stars one day post-creation; (3) independent April articles (Awesome Agents, others) attribute "roughly 1,000 GitHub stars within hours" - they pin it to the HN post, but the timeline shows the X tweet preceded HN and carried far more engagement, so the media's HN attribution is incomplete; (4) the star curve is sharply front-loaded: 377 (day 1) to 1.2k (day ~32, Apr 12) to 1,407 (day ~96, Jun 15), i.e. the vast majority of momentum was the launch-day spike, then a long slow plateau. SIGNAL: GitHub stars + founder-X engagement + HN points + downstream media all converge on the same 24-72h window, with the founder's X tweet as the leading edge. CONFIDENCE: High that the launch-day spike is the inflection and that the founder's X account, not HN, was the primary triggering channel (the X tweet's 1,774 likes vs HN's 47 points, and the chronology, are decisive). The CAUSAL driver underneath was the content itself: a concrete, credibility-borrowing hook ("autoresearch for GPU kernels," tied to Karpathy) plus immediately testable benchmark claims and a one-command quick-start, on a repo that was already polished (README, changelog discipline, Discord, progress chart) at first publication. Honest caveat: this is still a NICHE inflection in absolute terms (single thousands of stars; the company account itself has only ~1K followers), but the founder's *personal* X reach is materially larger than previously assumed - his openfang OS tweet hit 4,322 likes / 5,369 bookmarks - so the org's real megaphone is @Akashi203, not @rightnowai_co. The founders' later *standalone* HN promo (May arXiv + Show HN) flopped (4 and 2 points), but that does not contradict the X-led launch; it just shows HN was never their channel. The team treated AutoKernel as a launch-and-coast artifact (last push 8 days after launch) funneling credibility toward the commercial products (Forge / RunInfra).

Techniques & tactics

  • Borrowed-credibility framing: explicitly positioned as "autoresearch for GPU kernels," riding Andrej Karpathy's autoresearch concept (named in README headline, HN title, and credits).
  • Outcome-first, benefit-led KSP: "go to sleep, wake up to optimized Triton kernels": sells the result (overnight, no expertise) before the mechanism.
  • Concrete, falsifiable benchmarks up front (5.29x RMSNorm, 2.82x softmax, 2.21x cross-entropy on H100; FP4 matmul beating CUTLASS): gave technical readers something to argue about, which fueled the HN thread.
  • Launch where the audience lives: Hacker News first (correct venue for OSS dev tools), plus a same-week NVIDIA Developer Forums cross-post.
  • Ship a polished repo as the landing page: clean README, one-command uv quick-start, ships with self-contained example models (GPT-2/LLaMA/BERT, no transformers dependency needed), progress.png visualization, Discord badge, and a CHANGELOG from day one.
  • Standards/benchmark integration for legitimacy: built-in KernelBench (Stanford) integration and HuggingFace Kernels export: plugs into existing evaluation ecosystems rather than inventing its own.
  • MIT license + zero-friction adoption to maximize stars/forks.
  • Academic legitimacy layer: arXiv paper (cs.LG/cs.PF) + HuggingFace Papers + ResearchGate, ~11 days after the code launch, to seed a second media wave and signal seriousness.
  • Serial OSS portfolio strategy: AutoKernel is one of many free, attention-grabbing repos from the same org (openfang, picolm, etc.), each a top-of-funnel asset pointing back to the paid Forge/RunInfra products and the YC-backed company brand.
  • Founder-account-as-megaphone: the actual launch reach came from founder @Akashi203's personal X tweet (1,774 likes), amplified by the company account via retweet, not from an original company post. The founders did stay out of the HN thread (which read as organic third-party discovery), but the launch was not purely community-carried; the founder's own X post was the leading distribution event.

Sources