DEV Community: Simon Paxton

Ford Brought Back 350 Veteran Engineers After Its AI Quality Systems Missed Real-world Problems

Simon Paxton — Sun, 28 Jun 2026 19:33:22 +0000

Ford brought back about 350 veteran “gray beard” engineers over the last three years after its automated quality systems and AI tools missed engineering problems that experienced humans could catch. Those engineers were brought in to train younger staff, inspect vehicles and processes, and help reprogram the AI systems so the software would stop overlooking real defects.

That is the core answer here: Ford did not abandon automation, but it found that automation alone was not enough. In a factory, bad judgment is expensive in a way that looks a lot like the same failure mode seen in software tools when performance slips under reduced scrutiny, as in this case of AI quality dropping after reduced reasoning effort.

Ford brought back 350 veteran engineers after automated quality systems missed problems

The clearest reported figure is 350 rehired veteran engineers, cited by Bloomberg Law and summarized by TechCrunch. The span matters: that number covers roughly the last three years, not one sudden rehiring burst.

Bloomberg Law reported that Ford executives said AI-driven and automated quality systems were missing issues in engineering and manufacturing work that veteran employees could identify from experience. In plain terms, the machines were seeing the checklist, but not always the failure.

That distinction matters in automotive manufacturing because many defects are not cleanly labeled or easy to reduce to a sensor threshold. Some are pattern-recognition problems shaped by context: the odd noise, the part that technically fits but wears badly, the assembly variation that becomes a warranty claim months later. Ford’s move is a concrete reminder that AI accountability still depends on human oversight when the system touches the physical world.

Ford paired veteran know-how with AI retraining and junior-staff mentoring

Ford did not bring these engineers back just to stand at the end of a line and point at flaws. According to Bloomberg Law’s reporting, the veterans were needed for three jobs that AI could not do well on its own:

Teach junior engineers what failure looks like in practice
Catch missed quality problems in products and production
Retrain and reprogram AI tools so the systems learn from better judgments

That third point is the important one. If an automated quality system misses the wrong things, it does not just need more compute or more cameras. It needs better examples, better thresholds, and better judgment embedded in the training and review loop. The veteran engineer becomes part inspector, part teacher, and part source of ground truth.

This is a familiar pattern in AI deployment. Systems often work best not when they replace experts, but when they absorb expert feedback over time. In Ford’s case, the missing ingredient appears to have been tacit knowledge—the kind of know-how that is rarely captured fully in documentation but shows up in years of seeing parts fail, tolerances drift, and field complaints repeat.

Ford itself has publicly framed quality as a major operational focus. In a July 2025 quality update, the company said it was making progress on launch quality and warranty trends while also acknowledging that recalls remained a serious problem. The company’s investor-hosted PDF version of that update made the same basic point in more formal terms: quality was improving in some measures, but product safety and recall work were still unfinished.

Ford’s quality rebound coincided with recall pressure and a No. 1 mainstream J.D. Power finish

The best public evidence that the shift helped is Ford’s result in the 2026 J.D. Power U.S. Initial Quality Study. Ford ranked No. 1 among mainstream brands in that study, which surveyed 78,514 purchasers and lessees of new 2026-model-year vehicles after 90 days of ownership.

The study measures problems per 100 vehicles in the first three months. In the official 2026 PDF, Ford posted 208 problems per 100 vehicles, better than the mainstream-brand average of 212.

Measure	2026 result
Ford problems per 100 vehicles	208
Mainstream brand average	212
Owners surveyed	78,514
Ownership window measured	First 90 days

That is a real result, but it is not the whole story. J.D. Power’s Initial Quality Study measures owner-reported issues in the first 90 days, not long-term durability. A car can score well there and still contribute to larger recall trouble later.

Ford’s broader recall history remained a major issue even as initial-quality numbers improved. Ford’s 2026 Integrated Sustainability and Financial Report shows the company was still dealing with elevated recall and product-safety pressure into 2025 and early 2026. So the right conclusion is narrower than “AI failed” or “Ford fixed quality.” The stronger conclusion is that experienced engineers were still necessary to make the automated system useful.

That is also the more interesting lesson for manufacturing. AI is good at scaling repeatable checks. Veteran engineers are good at noticing when the check itself is wrong. In a car plant, that difference is the difference between a dashboard metric and a field repair.

The next public checkpoint will likely be Ford’s future quality and recall disclosures, including whether its internal improvements continue to show up in company quality updates and external studies like J.D. Power’s annual IQS.

Key Takeaways

Ford brought back about 350 veteran engineers over the last three years because its automated quality systems and AI tools were missing real engineering problems.
Those veterans were reportedly used to inspect quality issues, mentor junior staff, and help retrain the AI tools.
Ford later finished No. 1 among mainstream brands in the 2026 J.D. Power Initial Quality Study.
The J.D. Power study covered 78,514 owners and measured problems in the first 90 days, not long-term reliability.
Ford’s improved initial-quality ranking did not erase the company’s broader recall pressure in 2025 and early 2026.

Notion Mail Is Shutting Down 14 Months After Launch, and It Shows Which Software Layers AI Agents Eat First

Simon Paxton — Thu, 25 Jun 2026 19:38:32 +0000

Notion Mail is shutting down, and its strongest implication is that inbox wrappers, scheduling assistants, and workflow layers that mostly organize routine work are the first software categories AI agents replace. The key supporting fact is Notion’s own claim, reported by TechCrunch, that more than half of Notion Mail users already manage email without opening the inbox.

The timeline is unusually short. Notion formally launched Notion Mail on April 15, 2025, after previewing it in late 2024 as an AI-powered Gmail client, and announced its shutdown on June 25, 2026 — roughly 14 months after release.

Notion Mail’s shutdown timeline and product pivot

Notion Mail was not a replacement for email itself. It sat on top of Gmail, adding AI-written drafts, views, scheduling, snippets, and inbox organization for people who wanted a smarter front end to an existing account. That matters because the shutdown removes a layer, not the underlying mailbox.

Notion’s strategy changed fast. The company introduced Custom Agents in February 2026 as tools for repetitive work, then launched its Developer Platform on May 13, 2026 with Workers, synced data, and custom tools for outside systems. The product line moved from “better interface” to “software that acts.”

That makes the shutdown look less like a one-off product miss and more like a pivot away from UI-heavy wrappers. In its June 25, 2026 shutdown report, TechCrunch says Notion plans to keep agent-based email workflows alive even as the standalone inbox goes away.

There is a precedent here, too. Notion acquired Skiff in February 2024, then Skiff’s service sunset on August 9, 2024, with forwarding ending on February 9, 2025. Notion’s email push was real, but so was its willingness to collapse products into a different stack when the strategy changed.

The software layers AI agents replace first

The clearest pattern is that agents first eat software that mainly triages, routes, reformats, schedules, or summarizes work that already exists somewhere else.

A simple way to see it:

Category	Why agents replace it early
Inbox wrappers	Email can be sorted, drafted, labeled, and replied to without opening a custom inbox UI
Scheduling assistants	Calendar coordination is repetitive, rule-following work with clear constraints
Workflow layers	Many tools mostly move information between apps, assign status, and trigger routine actions
Agent-orchestration layers	These can survive, but only if they become the layer agents use rather than a human-facing dashboard

Notion Mail fits the first bucket. Its original pitch was an inbox that “thinks like you”, with personalized organization and AI help for handling messages. But once an agent can read a message, compare it with your past behavior, draft a response, and file it correctly, the inbox itself starts to look like optional scaffolding.

That is the same broader shift behind Google’s own Gmail AI personalization shift: the value moves from a visible interface to invisible decision-making. If the software can act on your behalf, the interface becomes something you open for exceptions, not for routine flow.

Scheduling tools are vulnerable for the same reason. A scheduling assistant is often a thin layer over rules: preferred times, participants, availability, follow-ups, and reminders. That is agent food. The same goes for workflow software that mostly exists to pass tasks from one queue to another, update statuses, or trigger standard actions after a document lands in a system.

Much of the public evidence here still comes from vendor product launches, platform strategy documents, and product announcements rather than broad market usage studies. But the direction is consistent: the first things to go are the layers whose main job is routine coordination.

Why inbox apps lose value when agents can act without a UI

Inbox products are especially exposed because the inbox is often a dashboard for decisions that can be made elsewhere. If an agent can decide archive, label, draft, escalate, schedule, or ignore, the main reason to open a dedicated email client disappears.

Notion’s own reported usage pattern is the blunt version of that thesis: more than half of users managed email without opening the inbox. That is not an independently published dataset, and Notion has not published Notion Mail user counts, revenue, or retention. But it is still the most important fact in the story because it describes the product becoming unnecessary in the exact place it was supposed to matter.

“More than half of Notion Mail users already manage email without opening the inbox,” TechCrunch reported from Notion’s rationale for the shutdown.

Competing tools are reacting in two different ways. Superhuman is trying to survive by becoming controllable by outside agents, including Claude and ChatGPT, instead of insisting the user live inside its own interface. Fyxer is positioned even more directly as a background assistant inside Gmail and Outlook, not as a new inbox you must switch to.

That is the implication for the category: products that are just a nicer inbox are in trouble; products that become the execution layer for agents may survive. The same pattern is showing up in coding tools, where the winners are not merely prettier editors but systems that can take work, run it, and hand back results inside broader AI coding agent workflow leaders.

The surviving layer is not “email app.” It is closer to an agent control plane: permissions, actions, memory, tools, and auditability across existing systems like Gmail, calendar, docs, and CRM.

Notion’s own Developer Platform release points in exactly that direction. Workers, synced external data, and custom tools are infrastructure for agents to operate across systems. That is a very different business from persuading users to adopt one more inbox.

The next concrete fact is simple: Notion announced the shutdown on June 25, 2026, and its replacement strategy is agent-driven workflows rather than a second attempt at the same email client.

Key Takeaways

Notion Mail is shutting down, and the clearest lesson is that inbox wrappers, scheduling assistants, and routine workflow layers are early targets for AI agents.
Notion Mail launched on April 15, 2025 and was shut down on June 25, 2026, a life span of about 14 months.
Notion told TechCrunch that more than half of Notion Mail users managed email without opening the inbox, which is the strongest evidence for why a dedicated inbox UI lost value.
Superhuman and Fyxer suggest two survival paths: become agent-controllable or become a background assistant inside existing mail systems.
Notion’s Custom Agents and Developer Platform show the company is moving from user-facing organization software toward agent execution and orchestration.

A 0.42-Million-Year-Old Tooth Just Complicated Human Intelligence Evolution

Simon Paxton — Sun, 17 May 2026 21:37:07 +0000

A Nature paper published on 13 May reports that enamel proteins from six Chinese Homo erectus fossils are adding molecular evidence to the branching history behind human intelligence evolution, even if the study does not identify where intelligence itself came from.

The team analysed teeth from Zhoukoudian, Hexian and Sunjiadong in China, using ancient enamel proteins rather than DNA — which is usually long gone at this age — to place the fossils more precisely in the hominin family tree.

For years, the broad story of human origins was tidy enough for a textbook: Homo erectus spreads widely, later lineages branch off, and eventually the line leading to us acquires the cognitive traits bundled into human intelligence evolution. The new study keeps the fossils but makes the tree messier, which in this field is usually what better data does.

The Nature paper says the Chinese H. erectus record spans roughly 2.1 million to 1.6 million years ago and disappears around 0.4 million to 0.3 million years ago. One of the key specimens, a tooth labelled PA69 from Zhoukoudian, comes from layers dated to about 0.42 million years ago.

“Our study shows that paleoproteomics can provide molecular evidence for Middle Pleistocene hominins where ancient DNA is unavailable,” the authors wrote in the paper. That matters because enamel proteins survive where DNA often does not, giving researchers a way to test relationships that were previously argued mostly from skull shapes and jaw fragments.

Live Science reported that the results point to “deep genetic links” between H. erectus and modern humans, with the implications running through Denisovan-related ancestry. That does not mean these fossils reveal the origin of thought, language or planning. It means the ancestry of the populations that eventually produced those traits may have involved more overlap and more surviving branches than older linear models allowed.

The study looked at six individuals dated to about 400,000 years ago, drawn from three Chinese sites. The paper's contribution is partly technical: it combines fossil sampling with newer experimental and computational methods for reading ancient enamel proteins. In a field where the usual outcome is “the biomolecules are too degraded,” getting usable molecular data from this period is the interesting part.

An independent expert reaction in the available coverage was blunt. “It raises the question of whether we know what Homo erectus even is,” palaeoanthropologist John Hawks at the University of Wisconsin–Madison, who was not involved in the study, told Live Science. That is about taxonomy, but taxonomy is doing real work here: if the labels on these populations are fuzzier than scientists thought, then clean stories about the ancestry behind human intelligence evolution get fuzzier too.

Context from other recent fossil finds points in the same direction. Scientific American reported in January on 773,000-year-old Moroccan fossils that may sit near the ancestry of Homo sapiens, Neanderthals and Denisovans, while a Nature commentary argued those remains illuminate African ancestors of our species. The annoying habit of the evidence, if you like dry summaries, is that every new well-dated find seems to add another branch instead of removing one.

This was a protein study on a handful of fossils, not direct evidence about cognition, and the paper does not claim to explain intelligence.

The next milestone is likely more molecular sampling from additional Middle Pleistocene fossils, as researchers test whether enamel proteins can clarify where East Asian H. erectus fits relative to Denisovans and other archaic humans. The study appears in Nature as “Enamel proteins from six Homo erectus specimens across China.”

Key Takeaways

A Nature paper published on 13 May 2026 analysed enamel proteins from six Homo erectus fossils in China.
The fossils came from Zhoukoudian, Hexian and Sunjiadong, with one key Zhoukoudian tooth dated to about 0.42 million years ago.
The study uses ancient enamel proteins because DNA usually does not survive in fossils this old.
The findings complicate the hominin family tree linked to human intelligence evolution rather than directly explaining the origin of intelligence.
Outside commentary says the results may blur the taxonomic boundaries around Homo erectus itself.

DEV Community: Simon Paxton

Ford Brought Back 350 Veteran Engineers After Its AI Quality Systems Missed Real-world Problems

Ford brought back 350 veteran engineers after automated quality systems missed problems

Ford paired veteran know-how with AI retraining and junior-staff mentoring

Ford’s quality rebound coincided with recall pressure and a No. 1 mainstream J.D. Power finish

Key Takeaways

Further Reading

Notion Mail Is Shutting Down 14 Months After Launch, and It Shows Which Software Layers AI Agents Eat First

Notion Mail’s shutdown timeline and product pivot

The software layers AI agents replace first

Why inbox apps lose value when agents can act without a UI

Key Takeaways

Further Reading

A 0.42-Million-Year-Old Tooth Just Complicated Human Intelligence Evolution

Key Takeaways

Further Reading