llama.cpp ships MTP support, speeds Qwen3.6 by 78%
Today's top 18 insights for PM Builders, ranked by relevance from X, YouTube, Blogs, and LinkedIn.
llama.cpp ships MTP support, speeds Qwen3.6 by 78%
#1 đť•Ź
clem 🤗 – Co-founder & CEO @HuggingFace unveils llama.cpp’s new MTP support, delivering a 78% speed boost on Qwen3.6-27B dense generation (25→45 tok/s) on an A10G.
#2 ▶️
How This 5x Founder Runs His Startup Solo With AI Agents (OpenClaw, Codex, Devin) | Ryan Carson
Peter Yang
Ryan Carson demonstrates how he leverages OpenClaw's ClawChief cron jobs and markdown skills together with Codex and cloud-based Devin to automate his executive assistant workflow, nightly sales prospecting via the Firecrawl API, and ship over 10 pull requests per day.
- The “executive assistant sweep” cron in OpenClaw’s ClawChief setup runs every 15 minutes to check Gmail via the Google CLI, sync Todoist tasks, parse and book Calendly links, ping updates in Slack threads, and proactively follow up on emails.
- The nightly “prospecting skill” cron uses the Firecrawl API ($20/month) to scrape LinkedIn data for family law attorneys and mediators into a Google Sheet CRM, then drafts and sends cold outreach from Ryan’s ryancarson.com email.
- In a cloud-based Devin environment, Ryan runs scheduled automations with playbooks—such as weekly end-to-end user journey tests—and applies a “land PR” skill to review and merge code, resulting in at least 10 pull requests shipped per day.
#3 đź“ť Mario Zechner
Constraint Decay: The Fragility of LLM Agents in Backend Code Generation - Fixing a unified API contract across 80 greenfield generation tasks and 20 feature-implementation tasks spanning eight web frameworks and evaluating with end-to-end behavioral tests plus static verifiers reveals that as structural constraints accumulate agents lose on average 30 percentage points in assertion pass rates from baseline to fully specified tasks (with some weaker configurations approaching zero); agents succeed in minimal, explicit frameworks like Flask but perform substantially worse in convention-heavy frameworks such as FastAPI and Django, and most failures are caused by data-layer defects (incorrect query composition and ORM runtime violations).
#4 đť•Ź
Teresa Torres is co-developing an AI-driven customer interview tool (with a partner building the UI) that uses git-diff–style change sets to highlight what’s changed between rounds and a two-step synthesis process—first surfacing noteworthy points in each interview, then spot...
#5 đť•Ź
Dharmesh Shah argues that while AI models now excel at reasoning and large-context understanding, it’s the harness—platforms like ChatGPT or Claude Cowork that supply tools, memory, skills, and context—that truly turns a powerful model into a usable product.
#6 đť•Ź
Garry Tan – President & CEO @ycombinator argues that while most AI agent builders focus on the “prefrontal cortex” (planning and reasoning), true leverage comes from building the “cerebellum” that automates mundane, repetitive tasks.
#7 đť•Ź
clem 🤗 – Co-founder & CEO @HuggingFace 300,000 AI builders have filled out their hardware profiles on @huggingface, and we’re publishing the aggregated insights at huggingface.co/hardware. We can’t wait to see how these trends evolve with the surge in local AI.
#8 ▶️
The AI paradox: More automation, more humans, more work | Dan Shipper
Lennys Podcast
Dan Shipper describes Every’s custom “senior engineer benchmark” that asks models and engineers to rewrite their vibe-coded Proof application from first principles, showing GPT 5.5 (Opus 4.7 plan) scored 62/100 versus human engineers in the high 80s to low 90s.
- All coding models prior to GPT 5.5 scored 30/100 on the senior engineer benchmark.
- GPT 5.5 running on the Opus 4.7 plan achieved 62/100 on the benchmark rewrite.
- Human senior engineers each scored in the high 80s to low 90s out of 100 on the same benchmark.
#9 đť•Ź
Peter Yang raised a $2M seed round but is holding off on hiring so he can personally learn each role’s pain points. Instead, he’s onboarding AI agents for faster ramp-up and ongoing training improvements.
#10 đť•Ź
Madhu Guru argues that CEOs’ AI FOMO drives them to set broad, vague AI mandates without any hands-on leadership. This yields performative demos and years of stalled progress, leaving them ripe for disruption by nimbler startups.
#11 đť•Ź
Lenny Rachitsky shares Dan Shipper’s view that AI won’t trigger mass unemployment but will commoditize yesterday’s skills—real value comes when humans creatively recombine that “frozen competence” into new, unique solutions.
#12 đť•Ź
Logan Kilpatrick calls for shifting AI narratives away from technical specs toward the real-life benefits and opportunities it creates, urging everyone to tell that outcomes-focused story.
#13 đť•Ź
Logan Kilpatrick highlights that there are already thousands of little-known companies and products achieving positive margins and growth—and he expects that number to swell to millions soon.
#14 đť•Ź
Garry Tan – President & CEO @ycombinator: The best startups are built by founders with a specific, hard-earned insight from living inside a problem, not by generic “AI for X” plays.
#15 đź“ť Mario Zechner
Building Pi With Pi - Pi, now part of Earendil but still Mario’s project, is being used to build itself and the team reports that many LLM/clanker-produced issue reports are noisy, inaccurate, and lead to over-engineered code changes rather than fixing root causes. Over the past 90 days (excluding Earendil members) the public GitHub tracker received 3,145 external issues/PRs, 2,504 were auto-closed, 17% were reopened (rising to 26% if counting issues referenced by main-branch commits), and only about 8% of auto-closed PRs (60 of 714) were ultimately merged.
#16 đź“ť Armin Ronacher
Building Pi With Pi - Pi is used to build Pi, but the team finds LLM-assisted "clanker" output often produces inaccurate, overengineered issue reports that mislead agents despite a custom /is command instructing Pi not to trust issue analysis. Over the last 90 days (excluding Earendil members) the public GitHub tracker had 3,145 external issues and PRs, 2,504 were auto-closed, 17% were reopened (rising to 26% if counting issues referenced by main-branch commits), and 60 of 714 auto-closed PRs were ultimately merged (about 8%).
#17 đť•Ź
Yann LeCun – Professor at NYU & Executive Chairman at AMI Labs; Ex-Chief AI Scientist at Meta unveils a self-supervised “Predictive Sparse Memory” framework that fuses sparse latent encoders, momentum-based updates and energy-based losses to cut training compute 10× and halve...
#18 in
Dharmesh Shah emphasizes that AI platforms like ChatGPT and Claude Cowork—providing tools, memory, skills and context—matter far more than the underlying model alone.