David Flickinger

Your best answers often already exist in old proposals, project folders, inboxes, and a couple people's heads. Vellm Lite is one agent connected to all of it. You own it, we run it.

For my first post, I’m sharing a letter @NVIDIA signed on why open models matter. AI will transform every industry, power every company, and be built by every country. Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. The world needs both frontier closed models and frontier open models. images.nvidia.com/pdf/Open-Weigh…

@sofiafern964 @PrivatEquityGuy Easier to implement than large enterprise, and immediate measurable impact

@DWFlickinger @PrivatEquityGuy smb ai is the sweet spot

Always great to chat with @PrivatEquityGuy on the latest AI trends across the SMB landscape

My conversation with @DWFlickinger on how business owners, private equity professionals, and CEOs can move beyond using Claude and ChatGPT as simple productivity tools and begin building AI infrastructure that creates real enterprise value. “This $6.5M project estimate took AI 23 minutes to generate vs. one of our estimators over a week...the difference was $400. More proposals sent and more jobs won without hiring additional staff.” I told David that I expected this episode to be so practical on AI that everyone who listens and has a desire to double their $1-10M EBITDA business would immediately forward it to their co-founder, portco operator or CEO. I hope you enjoy it, learn a lot and most importantly, implement asap. Timestamps: 0:00 Introduction: Why Most Business Owners Use AI Wrong 1:20 David Flickinger's Journey from Marine Officer to AI Operator 6:48 ChatGPT/Claude vs Real AI Infrastructure 0:24 The Roofing Company AI Case Study 6:17 How AI Learns Decades of Business Experience 20:39 Human Oversight, Trust & AI Decision-Making 25:30 What Happens to Junior Employees in an AI World? 28:07 How AI Doubled Revenue Without Hiring More Staff 32:17 AI, Key-Man Risk & Higher Business Valuations 39:02 The Biggest Risks of Implementing AI 45:05 The First AI Project Every Business Owner Should Start

Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5. This is a 17-place jump from Kimi-k2.6 (#18 -> #1). In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2 only in Gaming behind Fable 5. The full model weights will be released by July 27. Congrats to the @Kimi_Moonshot team on this major milestone!

@Teknium So sorry man. They become such a part of our lives it’s heartbreaking when they pass…Deepest condolences.

Hey everyone. I haven't been very responsive on here the last week. My dog, Link, who I've raised since he was a puppy over the last 13 years, passed away yesterday after being in the vet ER's ICU since last Wednesday for heart failure. I put together some of my favorite pics of him to share so you all can see the most awesome animal friend I could ask for. I'll be a bit slow probably through this week too, hope you all can understand 🙏

We have seen this to be true in lower middle market. Hiring is easier to justify if you know you’re getting more production. Integrate AI, grow faster, hire more.

@AlexFinn Can’t imagine running a small business with any aspirations of “using AI” and not having a plan to integrate local models in the very near future.

Fable 5 will be API pricing starting July 7th. For the next week, you can use up to 50% of your usage on Fable This will use up your limits spectacularly quick The days of subscription subsidies are quickly coming to an end. There will come a day soon where subscriptions will no longer exist for any service. Everything will be billed on usage All of these companies are trying to IPO in the next year. They need to become profitable. Right now they are all radically unprofitable The world is changing. It's time you start preparing for a world where you pay for every single token you use (unless of course, you're using local AI)

Coding is just first workflow where the inputs, outputs, and feedback were most well defined...makes sense everyone sees it there...but margin (in real economy) comes when agents automate real ops: bids, schedules, invoices, service requests/customer follow-up. etc. Can't speak for S&P, cause they're aren't our customers, but AI return will definitely show first as EV in private owner-led companies, well before SP 493. We are seeing really impressive productivity gains across lower middle market!

This is really important. AI is absolutely great for coding. But right now I don't see a lot of big use cases with returns in the real economy...

@anita_joshii Absolutely! There will be a big difference in outcomes…those who have vs those who don’t

@DWFlickinger operating layer matters.

The private-capital AI question is shifting. 2024: “Which model should we use?” 2025: “Which AI tools should our companies buy/build?” 2026+: “Who owns the operating layer?” The model matters, but it is not the strategy. In a real operating business, advantage compounds in the layer that knows: - how the workflow actually runs - which exceptions matter - where judgment enters - what the best operator checks - which systems must stay in sync - what changed after the last correction That layer should not live inside a random collection of SaaS tools, chat histories, one-off automations, or a single model provider’s ecosystem. It should be client-owned infrastructure. The best AI deployments I’m seeing start with one workflow wedge, not a company-wide AI mandate. Pick a workflow where judgment, documents, systems, and delay collide. Install the operating layer there. Then let every correction, exception, and decision make the next workflow easier.

@theshaneemoret @pmarca Well, despite the recent regulatory disappointment, I’ve been incredibly encouraged by the open source community. It may not be as obvious or easy as downloading a desktop app, but I’m hopeful those projects will give SMB owners viable options regardless of what governments do.

@DWFlickinger @pmarca I said the same in a post a few hours ago. These models give SMBs leverage. Now they are being gatekept from SMBs? Interesting…

More and more people are saying.

Loop Engineering is the next step after prompt engineering. Most people still use Claude Code, Codex, Cursor, or Grok like a chat box: Prompt. Wait. Copy. Fix. Prompt again. This repo shows the next step: You stop prompting the agent. You design the loop that prompts the agent for you. Inside: → Daily triage loops → PR babysitter loops → CI sweeper loops → Dependency sweeper loops → Changelog drafter loops → Post-merge cleanup loops → Issue triage loops It also gives you CLIs to: • Scaffold a loop • Estimate token cost • Audit if your repo is ready • Add memory/state • Add human handoff • Add verification gates • Run agents safely through GitHub Actions The wild part is the shift in thinking. Prompt engineering was about writing better instructions. Loop engineering is about building a system where agents keep working, checking, fixing, and escalating without you babysitting every step. This is what AI coding looks like when it stops being a chat session and starts becoming an operating system for software teams. Repo: github.com/cobusgreyling/…

Multi-agents collaborations are among the most interesting agent behaviors right now! We did an experiment the other day with 100+ agents (an open-collaborations for a week) collaborating to improve the inference speed of Gemma 4 in vLLM. Got a 5x final improvement in speed but what really stuck me was the interactions we observed on the message board Integrity & self-policing: - Social-engineering attempt: A human (FusionCow) asked agents to move to Telegram. An agent replied with an unprompted long post on "communication norms" refusing that, calling private side-channels "indistinguishable from collusion." - Verification loophole flagged: an agent found a relaxed verification loophole pushing TPS with clean PPL (PPL is teacher-forced, blind to decode divergence) and flagged it for a ruling by the community. The community pinged the human organizer which ruled it invalid. - Self-notice of overfitting risk: Some later improvements rested on pruning lm_head to a keep-set built from public PPL truth + public decode tokens. An agent noted this would lead to private-subset degradation and another built a keep-set explicitly covering eval prompts. Emergent collaborations: - Communal knowledge base: agents maintained shared lever-maps, playbooks, and triage tools so newcomers wouldn't repeat dead ends (stack-notes, playbook, int4-ceiling notes, MTP map, significance tool, policy simulator). - Four-agent relay: an agent built an int4-lm_head checkpoint but had no quota to run it; another agent tried to run it but failed at load, yet another agent diagnosed the config bug (tie_word_embeddings + ignore-list ordering) and a fourth agent was able to re-run and get to 118 TPS, 2.68×. Build/run/diagnose/ship ended up being split across four independent agents. - GPU-rich/GPU-poor division of labor: an agent was regularly compute-starved and switched to writing specs, byte-math, and acceptance analysis for other GPU-rich agents to execute. Some agents offered external Modal compute for another agent blocked DFlash training. - Cross-agent kernel debugging: an agent debugged another agent run of of yet another agent fused drafter: found a Triton store/load aliasing race in _k_qnorm_rope, a second shape bug, then rewrote attention with flash-decoding split-KV. Fixes posted "take freely." - Quota-pooling norm: Often agents would stage a candidate publicly for whoever has quota to run it. Agents will then usually credits the originator. This behavior emerged because of the 10-job/24h cap (e.g. pupa's package run by resystagent and fabulous-frenzy). Discoveries & reversals: - Agents would make many discoveries and reversal of them, giving them names like the following: - 127 TPS "wall" was an artifact. a mathematical proof of the max possible speed became called in the community the "int4-Marlin floor" but a later agent called the proof circular (only varied the bandwidth term, never overhead). Finally another agent broke to 247 TPS via MTP speculative decoding on a vLLM nightly. - "Smarter draft loses." An agent showed that a 2B drafter's ~1 GB/token read dominates even at perfect acceptance and a much smaller 256-hidden drafter wins at batch-1 because its weights are nearly free to read. Agent discussed how per-accepted-token cost ≈ draft bytes read / acceptance. - "DFlash near-random acceptance": an agent remotly diagnosed the 2–5% acceptance rate of another agent as near-random, ruling out undertraining/vocab caps and pointing to a train/serve hidden-state mismatch (bf16 E4B extraction vs int4 serving). - Much of the race was noise: one agent decide to run the #1 submission 4 times and found a σ≈1.16 TPS variation in single run. Another agent confirmed across 358 runs / 66 buckets: frontier deltas <~4 TPS are ties. Community adopted a significance norm. So many interesting interactions in the interaction board: huggingface.co/spaces/gemma-c… You can explore also the lineage of inventions from the agents at: …gemma-fast-challenges.static.hf.space/index.html And the challenge it-self at …ma-challenge-gemma-dashboard.hf.space And the organization behind the challenge at huggingface.co/gemma-challenge

If you run a small business you’ll likely never get access to the best models. HOWEVER… If you connect your business to the right agent orchestrator you can extract just as many benefits, if not outcompete. Using Claude or GPT off the shelf when your competition always has one model better is a losing AI strategy!

@drewfallon12 And this should be validating. Great concept, and any business owner would be insane see the events of this past week and conclude going all-in on Anthropic is the right choice. Model agnostic tools seems like the only reasonable choice.

lol wasn’t there a startup that raised a ton of money to do this like a month ago

@prasenx All the more reason to have an agentic harness with fallback to GPT 5.5 or other models. You simply cannot have an AI strategy that is tied to one provider.

claude is down, are we getting fable 5 back or what

@JankDankins_ @Teknium @NousResearch We can set this up for you, but you should have strong justification and a budget. Can def compliment AR models in Hermes framework, but personally if you cannot setup + manage this yourself I wouldn’t recommend unless clear business need. Can get weird!

Hey @Teknium or @NousResearch, can you please add a diffusion parser so i can use DiffusionGemma on hermes please? I know its a lot to ask for a single model and am working on it myself but you have way more knowledge on how to fix it than i do potentially. (all my info has come from gemini and deepseek)

@BoringBiz_ Must have an architecture that outsources coding tasks to extremely low cost open-source agents. Very rarely does your code require that Opus 4.8 writes it.

This is coming to every company that got drunk on the AI hype and gave unlimited access to tokens for their employees Lot of them crunching the numbers and now just realizing how expensive tokens actually are

@Teknium The problem is they’re all good updates😅…so I actually do have to set my alarm!

It might be hermes update time be sure to set your alarm