Cultivating Tech Innovation | Advisor | Nashville Post one of 25 leaders who shaped Nashville over the past 25 years. #1 user of CareCurrent
Nashville / Global
Joined April 2011
- Tweets28K
- Following3.5K
- Followers27.7K
- Likes6.3K
Pinned Tweet
Full piece on the blog. Includes the cost breakdown, the tools I use, what I'd do differently, and why I think this is a generalizable approach for any builder with a real problem to solve.
carecurrent.app/blog/the-com…
Entrepreneur ran a story worth sitting with. A Marine Corps vet, Hannah Pollack, made an ice cream sandwich from scratch, put it on a restaurant's menu, and started her company with $20 and no business plan. Nightingale now does around $3M a month.
The $20 is the headline. The real lesson is the sequence. She made something people wanted before she built anything resembling a business. No deck, no funnel, no five-year plan. Just a product good enough that demand pulled the rest into existence.
I meet founders constantly who have the order reversed: elaborate plans, polished branding, and nothing anyone has actually asked to buy. In an era where you can generate all the trappings of a company in an afternoon, the discipline of making one thing people genuinely want is rarer, and more valuable, than ever.
Start with the thing people love. The business follows.
via Entrepreneur: entrepreneur.com/side-hustle
Carta's latest data shows the venture middle vanishing. H1 2026 vs H1 2025:
Seed nearly halved, $6.5B to $3.8B. Series A flat. Series B down. But Series C and later jumped from $23.9B to $31.8B. Total funding is up, yet the money is stacking at the top.
Same story by sector. Hardware nearly doubled. SaaS held. Everything in between got squeezed, with medical devices collapsing from $1.7B to $585M.
I have lived through two of these cycles. When capital concentrates like this, it is not caution. It is conviction chasing a small set of perceived winners while everyone else fights for scraps. Great companies still get built in the squeezed middle. They just have to prove more with less, earlier.
Raising below Series C right now? Plan for a longer, leaner road. The checkbook has moved upstream.
via Carta: carta.com/data/
OpenAI shipped GPT-5.6 and a new agent called ChatGPT Work on July 9. Three model tiers (Sol, Terra, Luna) and an agent built to finish whole jobs across your apps and files, not just answer questions.
I have watched this movie before. A platform starts as a helpful tool, then quietly expands until it wants to own the entire workflow. That is the play here. ChatGPT Work aims squarely at the layer where output actually gets produced: docs, spreadsheets, decks.
Two things I am watching. First, whether "finishes the job" survives contact with real enterprise messiness, or whether it dazzles in the demo and disappoints in the wild. Second, how much of your business you are willing to route through one vendor's agent.
The capability is real and moving fast. The lock-in question is the one leaders should ask now, before the workflows get entrenched.
via Axios: axios.com/2026/07/09/ai-open…
Meta just priced its new Muse Spark 1.1 API at roughly a quarter of what OpenAI and Anthropic charge: $1.25 per million input tokens against their $5. Zuckerberg called rival pricing "very extreme" and made undercutting the whole strategy.
I've watched this play run in every platform era. Once the leaders establish a category, a deep-pocketed competitor commoditizes the layer underneath them and competes on price. Meta can afford to run the model business near cost because the model was never the product. Reach is.
For anyone building on these APIs, the takeaway is practical: your inference costs are heading down, maybe sharply. Architect so you can switch providers without re-plumbing your whole stack. Loyalty to a single model vendor is a liability when the price floor keeps dropping.
Cheap is coming. Build to take advantage of it.
via CNBC: cnbc.com/2026/07/09/meta-jum…
New HBR piece makes a point every executive deploying AI needs tattooed somewhere visible: you outsourced the model, but you still own the risk.
When a system you licensed discriminates, mishandles data, or harms a customer, regulators and plaintiffs don't go looking for the vendor first. They look at the company that deployed it. That's you.
Thirty years in defense and healthcare technology taught me that accountability never transfers as cleanly as a contract implies. The frameworks exist to manage this. NIST's AI Risk Management Framework and ISO 42001 aren't box-checking exercises. They're how you prove you exercised diligence when someone eventually asks.
Demand vendor transparency. Plan for model portability so you're not trapped. And govern the AI you bought as seriously as the AI you'd build.
via Harvard Business Review: hbr.org/2026/07/you-outsourc…
Starbucks spends ~$400M a year on software. Now they're building their own to replace tools from Microsoft, IBM, and Oracle, targeting real budget savings this fiscal year.
For thirty years the build-vs-buy math was settled: buy, because building and maintaining custom software was too slow and too expensive for most companies. AI just rewrote that equation. Coding tools have made it fast enough for in-house engineers to compete with established platforms.
A word of caution from someone who's watched this movie before. Building is the easy part. Maintaining, securing, and supporting what you built is where the real cost lives, and it doesn't show up in year one.
If a $400M software line is making you rethink your vendor stack, good. Just budget for the maintenance bill that arrives later.
via Entrepreneur: entrepreneur.com/business-ne…
Ask Claude, ChatGPT, or Gemini for a random number between 1 and 10. You'll get 7. Almost always.
MIT Technology Review profiles Springboards, an Australian startup that built an LLM called Flint specifically trained to produce a wider variety of responses to open-ended questions. Most LLMs converge on the same answers because their training data and reinforcement tuning push them toward the statistically popular response.
Flint, built on top of Alibaba's open-source Qwen 3, identifies the specific moments in generation where variety matters — like before naming a destination when you ask "Where should I go in Europe?" — and increases randomness only at those points. The rest of the output stays coherent.
The groupthink problem has real consequences beyond vacation planning. When every AI tool recommends the same strategy, surfaces the same candidates, or suggests the same investment thesis, the diversity of thought that organizations depend on for good decision-making disappears. Quietly.
via MIT Technology Review: technologyreview.com/2026/07…
A new framework called MRAgent from the National University of Singapore resolves long-horizon queries using 118,000 tokens. LangMem needs 3.26 million for the same task. Twenty-seven times more.
The gap comes from a fundamental architectural choice. LangMem uses a static "retrieve-then-reason" approach. MRAgent lets the agent dynamically reconstruct its memory based on accumulating evidence, integrating the reconstruction into the reasoning process itself.
Runtime dropped in half too — from 1,122 seconds to 586 seconds compared to A-Mem.
Memory management will quietly become the competitive moat for production AI agents. The teams burning millions of tokens per query on memory retrieval face a cost problem that compounds with every additional user and every longer conversation. The teams who solve memory architecture early will own the economics.
via VentureBeat: venturebeat.com/orchestratio…
Faros AI analyzed two years of telemetry from 22,000 developers and 4,000+ teams. The headline numbers look great: task throughput per developer up 33.7%, PR merge rate up 16.2%.
Then the other numbers: incidents-to-PR ratio up 242.7%. Bugs per developer up 54%. Median review time increased 5X. And 31% more PRs merging without any review at all.
The relationship between AI adoption and defect rate doesn't flatten as organizations mature. The correlation strengthens as adoption deepens.
Faros calls the pattern "Acceleration Whiplash" — AI flooded a system built around human-paced development with output the system was never designed to absorb. The throughput looks like progress. The incident graphs tell a different story.
The companies treating AI code generation as a speed multiplier without upgrading their review and testing infrastructure will learn this the hard way.
via VentureBeat: venturebeat.com/orchestratio…
Carta published new compensation data that maps the AI talent war in hard numbers.
AI/ML engineer equity grants ballooned 31% between January 2024 and February 2026 — nearly triple the growth rate across the broader employee population. Salaries rose 9.1% over the same window.
The sharpest move came from smaller startups. At companies valued between $1M and $10M, AI engineer equity grants jumped 64% in two years. The $10M-$25M tier: up 52%.
At AI-native startups above $500M valuation, an 80th-95th percentile AI/ML engineer earns $320K salary plus 0.146% equity. At a non-AI startup with the same valuation, those benchmarks drop to $285K and 0.1%.
The talent market is pricing in something the org charts haven't caught up with yet: AI engineers are the new leverage point, and the companies competing hardest for them are the smallest ones with the most to gain.
via Carta: carta.com/data/AI-shifts-in-…
Alex Banks at The Signal digs into why smart people consistently underestimate AI. Two forces working together:
Amara's Law — we overestimate technology in the short run and underestimate it in the long run. And the "AI effect" coined by John McCarthy: once an AI capability works reliably, we stop calling it AI. We call it autocomplete, navigation, spam filtering, recommendations.
The tools disappear into the background the moment they work. Then we look around and say "real AI still hasn't arrived" while standing in a world already reshaped by it.
Banks argues this isn't innocent. The underestimation serves a purpose. If you acknowledge how much AI already does, you have to reckon with how much more it will do. Easier to call it "just a tool" and go back to business as usual.
via The Signal: thesignal.substack.com/p/you…
OpenAI and Broadcom built Jalapeño from initial design to manufacturing tape-out in nine months. Engineering samples are already running GPT-5.3-Codex-Spark workloads at production frequency.
The chip targets inference specifically — optimized around the kernels, memory movement, and serving patterns that matter most for frontier models. Performance per watt "substantially better" than current state of the art.
Google built TPUs. Amazon built Trainium and Inferentia. OpenAI just joined the custom-silicon club, and they did it faster than anyone expected.
The companies still renting Nvidia GPUs for inference at scale will feel the cost pressure first. The ones who built model-routing and multi-backend architecture won't need to panic.
via TechCrunch: techcrunch.com/2026/06/24/op…
Paul Baier documented two specific mistakes Claude Opus 4.8 made this week. Not hallucinations in an edge case. Straightforward factual errors a subject matter expert would have caught immediately.
The models keep getting better. “Better” and “reliable enough to trust without verification” remain different standards. The gap between those two is where expertise lives.
Every deployment I’ve seen succeed at scale shares one thing: a human who knows the domain catches the 3% the model gets wrong. That 3% is where the real damage hides.
via GAI Insights gaiinsights.substack.com/p/a…
Stanford’s DeLM framework cuts multi-agent AI costs 50% by killing the central orchestrator.
Most agent systems route every update through one coordinator. It dilutes findings, repeats failed paths, bottlenecks at scale. DeLM replaces it with a shared context that agents read and write directly.
Results: 50% cheaper on SWE-bench, outperforms GPT-5.4 and Claude Sonnet on multi-doc QA.
The orchestrator pattern felt elegant when agents were simple. At scale, the coordinator becomes the most expensive single point of failure. Stanford just showed you can cut it out.
via VentureBeat venturebeat.com/orchestratio…
MIT Tech Review profiled Subquadratic, a startup claiming to have solved the quadratic attention bottleneck in LLMs.
The numbers, if real: 56x faster than FlashAttention. 12M token context window. Running Opus 4.6 through RULER 128 costs ~$2,600. SubQ ran it for $8. Appen verified several claims independently.
The skeptics have a point — SubQ bootstrapped from Qwen weights, not trained from scratch, and sits behind a waitlist. One researcher called it “either the biggest breakthrough since the Transformer or AI Theranos.”
Independent validation moves this past the ignore-it stage. Worth watching.
via MIT Tech Review technologyreview.com/2026/06…
Benedict Evans' "AI eats the world," Spring 2026 edition is out — the clearest big-picture read on where this is heading. He also threw cold water on the charts ranking which jobs AI will "expose": they look rigorous and mostly aren't.
Thirty years in tech taught me the confident projections age worst.
via Benedict Evans ben-evans.com/presentations
MIT Tech Review: AI job panic lacks evidence, but entry-level roles are quietly disappearing. The real risk isn't automation replacing jobs. It's companies no longer creating the roles where junior workers learn.
via @techreview
technologyreview.com/2026/05…
Justin Welsh's best financial move was a job that paid $18k less, in a city he couldn't afford, with a résumé full of firings. Everyone reasonable said pass.
Reasonable choices compound quietly until you're in a life you never chose. Does yours still feel like yours?
via Justin Welsh justinwelsh.me/essays/unreas…
Every major platform now builds agents that act on your behalf inside existing tools. The question moved past “should I use agents?” to “which platform gives me the integration surface I need?”
via The Signal
thesignal.substack.com/p/goo…
The people who adapt fastest don’t have detailed 10-year plans. They have strong principles and short feedback loops. Good piece from Fortune on career strategy.
fortune.com/2026/05/21/shery…
