Home Our Work Expertise How We Work About Resources Join Our Team

Insights

Opinionated writing on AI, machine learning, and the business of technology, from our team. No hype, no paid placement. Just the view from 25+ years of doing this work.

32 posts

LatestAugust 13, 2026 · Rebecca Joy· LinkedIn

Entry-Level Hiring: Same Technology, Opposite Decision

Two companies, same quarter, same AI tools. One cut junior headcount; the other says AI is why it hired more. Rebecca digs into why the research and the boardroom are telling opposite stories. Harvard tracked 62 million workers across 285,000 firms and found junior employment down 7.7% at deliberate AI adopters, with no offsetting bump in promotions. Stanford's payroll data shows workers aged 22 to 25 in the most AI-exposed jobs down 16%, closer to 20% for young developers. Revelio Labs has entry-level postings down 35% since early 2023. Against that, a Strada survey of 1,498 executives says AI is growing entry-level hiring. Her read: the datasets measure different things. CVs and payroll capture what hiring managers do; surveys capture what executives believe. Calling a role more complex is not the same as filling it. The right management question isn't whether AI is good or bad for junior hiring, it's which dataset describes your firm and how recently you actually checked. And underneath the numbers sits the one nobody is asking: if you stop hiring juniors now, where do your seniors come from in ten years? Experience doesn't grow on demand. It never has.

Read on LinkedIn

July 30, 2026 · Rebecca Joy· LinkedIn

Hiring for the AI Era Is Not What You Think

Everyone says hire for curiosity and adaptability, but almost nobody says what that looks like across the table from a candidate. Rebecca argues the resume has quietly stopped being a signal: AI writes the CV, AI screens it before a human ever sees it, and hiring has arrived somewhere strange with the person filtered out of both ends. The process was broken long before AI — demanding one to two years of experience for the entry-level jobs that were supposed to create it — and now firms are pulling back on junior hiring the moment they expect the work to be automated, which means no one is planting the crop of future seniors. Her own approach skims the CV, reads the cover letter only for voice, then puts the paper down: unforeseen questions, situational questions tied to the job, and questions that push against a candidate's knowledge to see where teachability actually shows up. Every career is a data career now, real internships have to be part of the answer, and the fundamentals never changed. She isn't hiring a resume. She's hiring a person who can learn faster than the field moves.

Read on LinkedIn

July 16, 2026 · Rebecca Joy· LinkedIn

Who Do You Blame When the AI Is Wrong?

Deloitte's AI mistake cost them money and reputation, and the blame landed where it should: on the humans, not the model. But that case was easy. Rebecca asks what happens when the accountability chain has fifteen links, a product built on a model built on another model, and no one owns the whole thing. She reaches for the autonomous car problem, where programmers answered the life-and-death question years ago in a conference room, in code, without knowing anyone's name. Real harm is usually quieter and harder to trace: a recommender feeding content to a fragile mind, and a courtroom full of parties pointing at each other. Before the internet, a broken car had three possible culprits. Now responsibility just disappears. At Saigon A.I. the rule is human-in-the-loop from day one: if you didn't catch the mistake, the mistake is yours. Governance is the actual product, and tools like QCoda exist because the chain has to be visible before something goes wrong, not after. The Deloitte story isn't a cautionary tale about AI. It's a cautionary tale about leadership.

Read on LinkedIn

July 12, 2026 · Andrew Schwabe· Hackernoon

Lifestyle vs. Transactional Business Models, and the Very Different Cultures They Grow

Two founders, same office park, same headcount. One is building a home; the other is fixing up a house to sell. Andrew argues that a single unspoken choice, exit or succession, quietly reshapes everything inside a company: who you hire, how you treat a downturn, whether people are parts to be swapped or investments that compound. Neither model is the villain. The damage comes from running one while telling your team you're running the other. Two blunt questions reveal which game you're actually playing.

Read on Hackernoon

July 10, 2026 · Andrew Schwabe· Qcoda

The Right Model for the Right Task (And Why "Pick a Good Model" Is the Wrong Advice)

"Pick a good model" is incomplete advice. QCoda Labs Episode 3 runs a complexity ladder (XS/S/M task specs, ten trials per cell) and finds that two models are identical at low complexity, then diverge: qwen3-coder frays to 7/10 at M while gpt-oss holds 10/10. Neither is better. They fail at different points on the curve. Andrew also explains why the first inference broker was thrown out before it shipped, with four design errors, and why the right architecture is Measure, Gate, Rank, Escalate, built on per-execution attribution rather than a hand-authored dictionary that rots on day one.

Read on Qcoda

June 29, 2026 · Andrew Schwabe· Qcoda

We Tested Every Agentic Coding Engine Against Every Local Model. The Results Were Weird.

A structured bake-off: five coding engines against seven local models, ten trials each, same task. The best model did not win. A 6B flash model beat a 32B coder-tuned model when paired with the right engine, and aider's two modes produced opposite outcomes on the same model, 0/10 versus 10/10, purely on edit format. Andrew's conclusion: the "model is dropping my code" bug is an edit-format problem, not a model problem, and round-trip count is a first-class constraint on local hardware. Protocol compliance matters more than parameter count.

Read on Qcoda

July 2, 2026 · Rebecca Joy· LinkedIn

AI Won't Save a Broken Business Model

Bad data, broken processes, vague requirements and high expectations add up to one thing: a fast, expensive, broken result. Rebecca walks through the AI Project Kickoff Playbook Saigon A.I. built after having the same hard conversation with clients over and over: the five qualification questions that reveal whether a project is even ready, the discovery work that has to happen before anyone proposes a solution, the fit assessment most enthusiastic clients fail, and the twelve-item Go/No-Go checklist before a sprint gets booked. AI accelerates work that's already understood. It does not fix a business that isn't.

Read on LinkedIn

June 19, 2026 · Andrew Schwabe· Hackernoon

I'm Becoming a Progress Junkie (and AI is the Dealer)

AI makes you feel 20% faster. The research says you're actually 19% slower. Andrew traces the progress-junkie loop: the dopamine hit of constant motion, the cognitive offloading that quietly erodes your understanding, and the gap between feeling productive and being productive. The fix isn't more output, it's pacing, and knowing when the dealer is selling you a high instead of real progress.

Read on Hackernoon

June 19, 2026 · Andrew Schwabe· Qcoda

Your Coding Agent Is Silently Doing Nothing

Andrew spent an afternoon convinced he had a bad model. He didn't. Most qwen-family models return tool calls as plain text on raw Ollama endpoints, so agentic coding tools check for structured tool_calls, find nothing, report no edits, and exit cleanly with no error. The same model that goes 0/10 on one engine finishes in 21 seconds on another. vLLM without the right startup flag is worse: three identical requests at temperature zero return three different formats. The fix belongs at the serving layer, not in five per-client shims, and the default setup is broken silently.

Read on Qcoda

June 18, 2026 · Rebecca Joy· LinkedIn

Your Competitor Didn't Beat You With Better AI

They beat you with better clarity. AI amplifies whatever direction you're already moving, so if your strategy is fuzzy, it just moves you faster in the wrong direction. Rebecca compares today's AI rush to the late-'90s website scramble: the companies that rushed in without a plan paid twice, once to build it wrong and again to rebuild it. She's watched the same pattern repeat with AI: a green-lit tech team, six months of work, and a solution looking for a problem because the data wasn't ready and the use case was never defined. The hardest part is rarely the technology. It's getting the business side and the tech side in the same room, speaking the same language, long enough to build the same thing before a line of code gets written. That gap, alignment before automation, is the work behind both Saigon A.I. and QCoda.

Read on LinkedIn

June 11, 2026 · Andrew Schwabe· Hackernoon

What the Heck Is a Model Ladder? And Why Do I Need to Use One?

When you build agents on local, open-source LLMs, a model can simply disappear out from under you, and your agent grinds to a halt. Andrew's answer is the model ladder: an ordered-fallback pattern that tries your preferred model first, then climbs down to alternatives so the system keeps running instead of failing. He walks through the pattern in pseudocode and Python, and lays out the four decisions that make a ladder actually work, from how you rank models to when you give up and escalate. A practical playbook for anyone running AI in production who can't afford a single point of failure.

Read on Hackernoon

June 9, 2026 · Andrew Schwabe· Qcoda

Against the Giant

Frontier-lab AI charges you three coins: gold, secrets, and your own audit trail. Andrew's manifesto argues the third is the one that should keep a CTO awake, because you cannot audit a black box you rent by the hour. The myth that there is no third road is a marketing achievement, not a law of nature: open-weight models running on hardware you own handle the overwhelming share of what products actually ask of an LLM. The answer isn't a private frontier lab in your basement. It's matching each task to the right tool, with a record of why, and treating governance as the reins rather than the brake.

Read on Qcoda

June 4, 2026 · Rebecca Joy· LinkedIn

What AI Actually Costs

Everyone asks Rebecca about token costs. It's the wrong question. After one client's capped $200/month budget ballooned to $1,000 by month four, she lays out the real price of AI: the costs no vendor puts in a proposal. Data privacy that's a governance problem, not a technical one. The hunt for clean "Golden Data" she's never once seen in the wild. Change management that's underestimated every time. The people hours quietly absorbed into salaries. And the failed pilots nobody tracked well enough to even tally. The CFO ends up holding all the bills, and the question nobody wants to answer: was it worth it? The tool is not the investment, Rebecca argues. The transformation is, and it has a price that goes well beyond the token bill.

Read on LinkedIn

Want to go deeper on any of these topics?

We're always happy to discuss how these ideas apply to your specific situation.

Start a conversation