AI News Archive: June 16, 2026 — Part 9
Sourced from 500+ daily AI sources, scored by relevance.
- AI Mention Triage for Jira and Confluence
AI Mention Triage for Jira and Confluence Atlassian Marketplace
Score: 34🌐 MovesJun 16, 2026https://marketplace.atlassian.com/apps/485151906/ai-mention-triage-for-jira-and-confluence - Apple Stock Is Down After WWDC. There’s More to Its AI Strategy Than Meets the Eye.
Apple Stock Is Down After WWDC. There’s More to Its AI Strategy Than Meets the Eye. Barron's
- From camera clarity to collision avoidance: What ‘AI in mining’ actually looks like in 2026
From camera clarity to collision avoidance: What ‘AI in mining’ actually looks like in 2026 Mining Technology
- Why Fleet Has Both General Purpose Chat and Specialized Agents
Discussing the benefits of general purpose chat and specialized agents
Score: 34🌐 MovesJun 16, 2026https://blog.langchain.dev/blog/why-fleet-has-both-general-purpose-chat-and-specialized-agents - China's Tech Giants Deploy AI Across Every Front in the 618 Shopping Festival
Alibaba, JD.com, and ByteDance put AI at the center of China's 618 shopping festival, with AI assistants, digital hosts, and smart logistics reshaping e-commerce.
- Sloneek raises $6M to turn HR software into AI agents
HR SaaS startup Sloneek has raised $6 million in a newfunding round led primarily by Orbit Capital and Venture to Future Fund,significantly exceeding the company’s original target. The investment will...
Score: 34🌐 MovesJun 16, 2026https://tech.eu/2026/06/16/sloneek-raises-6m-to-turn-hr-software-into-ai-agents/ - How AI may be messing with home prices
For most real estate professionals, the data aggregation capabilities of AI can enhance their expertise. But there are shortcomings.
- Voice Cloning: Train AI to Write Exactly Like You — Prompt to Profit · Day 17 of 30
The next frontier isn’t getting AI to write well. It’s getting AI to write like you — so precisely that you can’t tell the difference. Here’s the exact system. There is a moment every writer knows. You read back a piece you wrote six months ago and something catches — a particular word choice, a rhythm in the sentences, the way a paragraph opens with a concrete scene before it earns its abstraction. You didn’t consciously put those things there. They emerged. They are yours. They are, in the truest sense, your voice. Now here is the interesting problem: most people’s AI-generated content sounds nothing like them. It sounds like a competent, confident, entirely characterless communicator — polished in the way that a hotel lobby is polished. Functional. Inoffensive. Forgettable. The reason isn’t that AI can’t write in a specific voice. The reason is that nobody gave it the blueprint. Voice cloning — training an AI model to write with your specific stylistic fingerprint — is one of the highest-leverage skills in this entire series. Done properly, it means you can produce a first draft at ten times the speed without sacrificing the element that makes your writing yours: the way it sounds. Today we cover exactly how to do this. The voice audit. The voice profile document. The sample calibration method. And the mirror test — the only reliable way to verify that your clone is actually working. Why Generic AI Writing Sounds Generic The default output of any AI model reflects the statistical centre of its training data — the most common patterns across billions of documents. That centre is not a bad writer. But it is a thoroughly averaged writer. No idiosyncrasies. No strong opinions about sentence length. No compulsion toward a particular kind of opening. No recurring structural moves that readers start to recognise as yours. Your voice exists precisely at the edges of that average. The places where you consistently deviate from the statistical mean — shorter sentences than most, em dashes where others would use commas, a habit of opening with a concrete scene before any abstraction — those deviations are your fingerprint. And AI, by default, irons them out. Voice cloning is the process of explicitly reintroducing those deviations. You’re not asking AI to invent a personality. You’re asking it to adopt a specific, documented set of stylistic constraints that match your existing patterns. The difference in output is immediate and significant. Same topic. Same AI model. Completely different result — because the second version had a voice profile loaded. The difference is entirely in the brief. Step One: The Voice Audit You cannot brief AI on a voice you haven’t clearly described yourself. The first step is a voice audit — a structured analysis of your existing writing that surfaces the patterns you repeat unconsciously. Gather five to eight pieces of your best work. Your favourite articles, your best emails, your most-shared posts. Paste them together into a single document and send them to AI with the following prompt: What you get back is your voice profile in raw form. Read it carefully. Mark the observations that feel true — that capture something you’ve felt but never articulated. Discard anything that seems like a generic observation that could apply to anyone. What remains is the foundation of your clone brief. Step Two: Building the Voice Profile Document The audit gives you raw material. The Voice Profile Document (VPD) turns that material into actionable instructions. Where the Master Memory Document from Day 16 tells AI who you are and what you’re building, the VPD tells AI how to write as you . They work together — load both at the start of any session where you need voice-matched output. Here is the template. Fill it with the specific findings from your audit. Be concrete — “direct” is not specific; “short declarative sentences, rarely more than fifteen words, with occasional single-word fragments for emphasis” is specific. Notice the final instruction — the mirror test built directly into the prompt. This is not rhetorical. It is a genuine instruction to the model: before you output this, check it against the profile. It produces measurably better first drafts than profiles that don’t include a self-check instruction. Step Three: Calibration and the Mirror Test You have a profile. Now calibrate it. Open a fresh session, load your VPD, and give AI a task you’ve done many times before — write a 300-word introduction on a topic you know well. Read the output against your actual writing. Not against a vague feeling of “that sounds like me.” Against a specific piece you wrote yourself. The calibration loop is systematic: What Changes When the Clone Works Once your voice profile is calibrated, the workflow changes fundamentally. You stop writing first drafts. You start editing first drafts — which is a different cognitive mode entirely. Faster, cleaner, less draining. The blank page problem disappears because there is no blank page. There is a draft that sounds like you, waiting to be sharpened. The realistic time saving for a writer with a well-calibrated voice profile is somewhere between 60 and 80 percent of drafting time. Not 100 percent — editing and judgment remain yours. But the effort of producing words on a page that sound like you? That part is largely solved. Your writing voice took years to develop. It deserves to be preserved and extended — not averaged out every time you open a new AI session. A well-built voice profile is, in that sense, a form of authorship protection. You’re not giving AI your voice. You’re giving AI the instructions to honour it. Build the profile carefully. Calibrate it honestly. And the next time someone reads a piece you produced with AI assistance, they should find it completely indistinguishable from the one you laboured over alone. That’s not a limitation on your craft. It’s a proof of it. Tomorrow, Day 18, we move to Prompt Economics — how to think about the true cost and value of every prompt you write, and the counterintuitive framework for deciding when AI is worth using and when doing it yourself is actually faster. For more resourcces and documents, please refer to the links in my profile page: Faheem Munshi — Medium Voice Cloning: Train AI to Write Exactly Like You — Prompt to Profit · Day 17 of 30 was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
- Insurers pivot AI strategy toward core risk underwriting
AI investments by insurers are now expected to generate tangible business value beyond mere efficiency. According to findings in the 2026 Evident AI Index, insurers are now embedding AI technologies into workflows that directly influence underwriting discipline and capital allocation. Christian Preece, Insurance Director at Evident, says: “For years, insurers have competed on AI ambition, […] The post Insurers pivot AI strategy toward core risk underwriting appeared first on AI News .
Score: 33🌐 MovesJun 16, 2026https://www.artificialintelligence-news.com/news/insurers-pivot-ai-strategy-toward-core-risk-underwriting/ - Phaidon International: Why the Future of Recruitment is AI-Powered but Human-Led
Phaidon International: Why the Future of Recruitment is AI-Powered but Human-Led USA Today
- Planview Closes the Gap Between Strategic Intent and Business Outcomes in the Age of AI
AUSTIN, TEXAS — Strategic portfolio management software provider Planview today introduces new capabilities for the future of AI-powered SPM. This includes the debut of the Outcome Intelligence Graph, a model of the enterprise portfolio in which every decision connects to the resources behind it and the results it produces. The Outcome Intelligence Graph addresses three … continue reading The post Planview Closes the Gap Between Strategic Intent and Business Outcomes in the Age of AI appeared first on SD Times .
- Karnataka explores biotech financing framework, AI-led innovation to support sector growth
Karnataka is developing a dedicated support framework for its biotechnology sector, encompassing funding, mentoring, and regulatory guidance to boost startups and growth capital access. Key focus areas for the next phase of growth include AI-enabled biotechnology and biomanufacturing, aiming to strengthen the state's innovation and investment ecosystem.
- French startup bets on non-humanoid design in crowded AI robot race
French startup bets on non-humanoid design in crowded AI robot race Reuters
Score: 33🌐 MovesJun 16, 2026https://www.reuters.com/business/french-startup-bets-non-humanoid-design-crowded-ai-robot-race-2026-06-16/ - Nvidia's Huang pledges AI will boost manufacturing jobs. A test will come in Texas
Nvidia is betting on artificial intelligence to revive U.S. manufacturing
Score: 33🌐 MovesJun 16, 2026https://abcnews.com/Technology/wireStory/nvidias-huang-pledges-ai-boost-manufacturing-jobs-test-133921258 - Lightbringer raises $10M to scale AI-powered patent services
Swedish legaltech startupLightbringer has raised $10 million in Series A funding to accelerate thedevelopment of its AI-powered patent platform and support its expansion intothe US market. The round w...
Score: 33🌐 MovesJun 16, 2026https://tech.eu/2026/06/16/lightbringer-raises-10m-to-scale-ai-powered-patent-services/ - How AI is identifying millions of plant species in 'biodiversity revolution' to boost conservation
How AI is identifying millions of plant species in 'biodiversity revolution' to boost conservation The National
- Probably raises $9M to build a more reliable kind of AI
Probably wants to prevent hallucinations and factual errors from reaching users, and achieve accuracy on par with deterministic systems.
Score: 32💰 MoneyJun 16, 2026https://techcrunch.com/2026/06/16/probably-raises-9m-to-build-a-more-reliable-kind-of-ai/ - What Separates Success From Failure in AI Implementation (Lessons from Automotive Retail)
Dek: AI requires responsibility and intention to use well. Here are examples of how automotive retailers are implementing AI effectively and safely. Remember the old Spider-Man quote? The one his Uncle Ben tells him (or his Aunt May in the new movies) about how “with great power comes great responsibility”? If there was ever a […]
- The Shift From SEO to AI Visibility: New Industry Insights Highlight Changes in How Businesses Are Discovered Online
The Shift From SEO to AI Visibility: New Industry Insights Highlight Changes in How Businesses Are Discovered Online USA Today
- IQanAI targets Gulf SMEs with AI sales agent platform ahead of September launch
iQanAI deploys AI sales agents that operate across WhatsApp, Instagram, Facebook Messenger, Telegram, SMS, iMessage, LINE, and web chat
- Run a Local LLM with OpenClaw on Your Mac Mini
Tired of your monthly API bill? Follow this tested guide to set up a high-performance local LLM on your Mac Mini without the headaches. The post Run a Local LLM with OpenClaw on Your Mac Mini appeared first on Towards Data Science .
Score: 32🌐 MovesJun 16, 2026https://towardsdatascience.com/run-a-local-llm-with-openclaw-on-your-mac-mini/ - Microsoft's Brad Smith on AI-era jobs: "Let's not panic"
Brad Smith , vice chair and president of Microsoft, slammed tech moguls for hypocritical, grandiose warnings that are alienating Americans at a time of huge workforce opportunity. "Nobody knows for sure, but let's not panic," Smith, who has been with Microsoft for 33 years , said from the tech giant's headquarters in Redmond, Washington. Why it matters: Smith is among the tech leaders who think dire predictions about AI's threat to entry-level white-collar jobs are souring young Americans on a miraculous technology. Smith writes in a new paper about AI's effect on jobs that AI getting booed at this spring's commencements should be a "powerful wake-up call for the tech sector." "Hopefully, leaders across our industry will listen and seek to learn from this reaction," Smith writes. "For the past half-century, the youngest generation of people and workers has led the way in adopting new digital technologies. ... When people who use a new technology complain about it, we had better take notice." In our interview, Smith said tech leaders have botched the conversation about AI and jobs: Hypocritical warnings: In an essay this month, Anthropic pointed to benefits of slowing AI development "to give ourselves more time to deal with its immense implications" — a so-called global pause . Smith told me: "If somebody says, 'This technology is so powerful that we need a global treaty to slow it down,' then I would say: Then take your foot off the accelerator yourself if you think it's moving too fast." Scaring grads: Echoing points Jim VandeHei and I made in our recent " Rattled Generation " column, Smith pointed out that this year's graduates were in high school during COVID and have done much of their socializing through screens, against a backdrop of political turmoil. "Now, they finally get to enter the workforce and here comes AI?" he said. "Too often, this is being presented to them as something that is going to happen to them, not for them. And I think this is their way of saying: 'Wait a second. Not so fast. We have a voice. We want to be heard.'" Short-term distortion: "This is going to unfold over 25 years, not two-and-a-half," Smith said. "But look, if you're trying to raise money as entrepreneurs need to do, it's easier to raise money if people think it's going to happen sooner rather than later. But we spend a lot of time studying the history of technology. Twenty-five years for complete transformation of an economy would be a record setter." Unrealistic hype: "Tech leaders tend to repeat two mistakes," he said. "One is: They overestimate the impact of technology, especially the pace at which it will arrive. And second: The tech leaders often underestimate people. ... There was a time when humanity discovered that a horse could run faster than a person. So, people learned how to ride horses. Let's use AI to help people do more and not replace us." Fake certainty: "You find that the same folks who made the wrong predictions a decade ago keep making them with extraordinary conviction," Smith said. "And it makes great fodder for people who generate stories for a living." Hollow calls for regulation: He said we're seeing a flashback to the past decade's debates about social media legislation. "You had some companies that said, 'We want legislation,' and then they basically opposed every bill in Congress because they never liked it specifically." He said that on AI policy, beware "ideas that are so grandiose that the chance of them being adopted is zero." The bottom line: The AI debate has been "too focused on grandiose predictions," Smith said, "and not centered enough in: Let's use technology, as we always have, to help people do better things." Share this video ... Read Brad Smith's post , "AI, jobs, and the next generation."
Score: 32🌐 MovesJun 16, 2026https://www.axios.com/2026/06/16/microsofts-brad-smith-on-ai-era-jobs-lets-not-panic - What is speech-to-speech for voice agents?
Exploring speech-to-speech technology for voice agents
- LocalMighty Reports Growing Shift Toward AI Search Optimization Among US Businesses
LocalMighty Reports Growing Shift Toward AI Search Optimization Among US Businesses USA Today
- How to Optimize Transformer-Based Models for Low-Precision Training
Transformer architectures are the backbone of many modern large language and generative AI models. As these models grow in size, training runs consume more GPU...
Score: 31🌐 MovesJun 16, 2026https://developer.nvidia.com/blog/how-to-optimize-transformer-based-models-for-low-precision-training/ - Galaxea AI Chief Says China Could Lead Robotics Models Within Three Years
Galaxea AI Chief Says China Could Lead Robotics Models Within Three Years Caixin Global
- AI Made Content Abundant. For Creators, Voice Is Now The Scarce Asset
AI made content abundant. A 2026 survey of 16,000 creators shows voice and judgment, not volume, are now the scarce asset
- AI-based system developed to better detect toxic online content
A Concordia-led team of researchers has developed a new artificial intelligence-based method of detecting toxic online content that is faster and more accurate than existing tools. The system is designed to ensure social media platforms can reliably prevent user-generated content they deem harmful from appearing online.
- Drilling Into AI’s Financial Sustainability
Budgets for AI tokens can’t be infinite, no matter how much hyperscalers wish they were The post Drilling Into AI’s Financial Sustainability appeared first on Towards Data Science .
Score: 30🌐 MovesJun 16, 2026https://towardsdatascience.com/drilling-into-ais-financial-sustainability/ - AI and the Art of Joyful Living
AI and the Art of Joyful Living Time Magazine
Score: 30🌐 MovesJun 16, 2026https://time.com/branded-content/igarden/ai-and-the-art-of-joyful-living/ - The AI Tool That’s Giving Landlords A Leg Up On Zillow
In the wise words of Spider-Man, with great power comes great responsibility. Nowhere is this more true, perhaps, than in the tech industry. In just a couple of years, consumers have become “AI native,” said Anda Gansca, CEO and co-founder of customer journey intelligence platform Knotch. This means that consumers have higher standards for the […] The post The AI Tool That’s Giving Landlords A Leg Up On Zillow appeared first on AdExchanger .
Score: 30🌐 MovesJun 16, 2026https://www.adexchanger.com/marketers/the-ai-tool-thats-giving-landlords-a-leg-up-on-zillow/ - The Sequence Knowledge #878: Beyond Transformer: What We Learned
And a new series about Distillation
- RBC chief executive turns to AI to help inform his daily agenda
Dave McKay spends at least an hour a day reviewing an AI-generated briefing on key issues he needs to know about
- Frontier post-training recipe review with Finbarr Timbers
"Interview" #18
- Agent Bricks: Data + AI Summit 2026
Last year at the Data + AI Summit, we launched Agent Bricks, ushering in a new way...
- Alibaba's AI Glasses Top China Sales Rankings at Just RMB 1,997
Alibaba has emerged as the early winner in China's rapidly growing AI glasses market, with its Tongyi Qianwen (Qwen) AI glasses ranking first in nationwide sale...
- How Hunar Is Using AI To Fix The Broken Economics Of Frontline Hiring
India’s AI conversation is increasingly shifting toward software agents, automation stacks and enterprise copilots. But amid the rush to automate…
Score: 30🌐 MovesJun 16, 2026https://inc42.com/startups/why-hunar-ai-thinks-indias-frontline-workforce-needs-ai-agents/ - How ISRO’s AI-powered coastal safety systems are helping detect dangerous rip currents in real time
Artificial intelligence is emerging as a powerful tool for coastal safety, helping detect dangerous rip currents through satellite imagery, weather data, ocean monitoring, and real-time analytics. AI-powered beach safety systems developed by ISRO can deliver faster warnings, support disaster management, improve public safety, and transform environmental data into actionable insights that help reduce drowning risks along coastlines.
- iFLYTEK Unveils AI Recorders: P1 and P1 Pro - Redefining Smart Voice Capture for Work and Study
iFLYTEK Unveils AI Recorders: P1 and P1 Pro - Redefining Smart Voice Capture for Work and Study USA Today
- Alibaba's Internal AI Agent Battle: QoderWork vs Wukong for Enterprise Crown
Alibaba's QoderWork outperforms DingTalk-based Wukong internally, as the company races Tencent and ByteDance to establish a flagship enterprise AI agent product.
- Infosys bags AI-led IT transformation mandate from Valmet
Infosys bags AI-led IT transformation mandate from Valmet Techcircle
Score: 29🌐 MovesJun 16, 2026https://www.techcircle.in/2026/06/16/infosys-bags-ai-led-it-transformation-mandate-from-valmet - Never Let the LLM Write the Joins
The subject graph, and the one architectural line that made me comfortable putting “talk to your database” in front of real users with real permissions. Part 3 of 4 on building a conversational analytics engine. ~11 min read. Part 2 built a map: three databases introspected and enriched into a domain graph. Tables, columns, join paths, business meaning, and the security rules, all sitting in a graph with a fast in-memory snapshot in front of it. A map answers nothing on its own. Nobody asks “describe the schema.” They ask “how many orders did Bike World place last quarter, in my territory,” and they expect a number, computed from data they are actually allowed to see. This article is the machine that does that. Back in Part 1, I listed seven walls that break text-to-SQL on real data. This is where four of them fall: the fan-out trap (Wall 3), ambiguity (Wall 4), security (Wall 6), reproducibility (Wall 7), plus the cross-source half of Wall 5. If Part 2 was the “why a map,” this is the “why a particular shape of pipeline.” And I want to lead with that shape, because it is the single most important idea in this entire series. The big idea: a query is not “hand the question to an LLM and let it call some tools.” I split it in two. Phase A is deterministic planning. It decides what to do (tables, entities, filters), scoped to who is asking. One LLM call, for reading intent. Everything else is code. Phase B is LLM-driven execution. It does the plan (writes SQL, runs it, recovers, answers). The line between them is the product. Security is never the model’s job. Joins are never the model’s job. The model does the creative parts; code does the parts that must be safe and reproducible. TL;DR The subject graph resolves the specific things a user names (“Bike World” becomes StoreID 620). It is a cache, not a copy. Phase A plans deterministically and produces a frozen plan. Phase B is a bounded ReAct loop that executes it. The LLM writes only the SELECT and WHERE. Joins come from the graph. Security is injected by rewriting the SQL. Cardinality is fixed by a planner. None of those are left to the model. Honest limit: today the conversation is single-shot . Follow-up memory is built but not wired. Throughout, one example: a sales rep I will call Dana , scoped to the Southwest territory, allowed to see sales data, asks “Show orders for customer Bike World.” Watch what each stage does to that sentence. The subject graph: concepts vs. actual things The domain graph from Part 2 knows about types . It knows CUSTOMER is a concept that lives in a table with a name column. It does not know that “Bike World” is a specific store with the key StoreID 620. That is an instance, and instances live in the subject graph . A canonical entity in the subject graph has: a stable id (something like entity:organization:bikeworld_a1b2c3d4) a name and type a set of aliases a list of source links , each pointing at exactly where the entity lives: StoreID 620, in the StoreID column, of Sales.Store, in the SQL Server source an access level , so resolved entities are protected by the same machinery as tables The defining property: it is a cache, not a copy. It is not pre-loaded with every customer. That would be copying the data, the one thing I refuse to do. It starts essentially empty and fills lazily , remembering only the names real users actually ask about. The rows stay in the source database. The subject graph just remembers the mapping from “a name a human typed” to “a key the database understands.” Resolving a name (the cascade) When the pipeline needs to turn “Bike World” into a key, it runs a cascade, ordered from cheapest-and-most-certain to most-expensive-and-least-certain. It stops as soon as it is confident. entity-resolution-cascade How to read it (with the real numbers, because the numbers are the design): Cache keyed on the name. Stores unfiltered results; RBAC applies after, so users with different permissions can share an entry. Exact match: canonical name scores 0.95 , an alias 0.90 . One clean hit above the 0.70 threshold short-circuits the whole thing. Fuzzy match: the rapidfuzz ratio scorer at a threshold of 80/100 , with tiered confidence so it never pretends to be as sure as an exact match ( 0.65 to 0.85 ). Semantic match: I will be straight with you. It is a placeholder. It sits in the right position in the cascade and returns nothing today. The plan is embedding similarity; the wiring is not done. Wherever you read “exact, fuzzy, semantic, source,” know the semantic rung is a stub. Source-DB lookup: hit the real table. An exact name scores 0.95 , and a success gets written back into the subject graph so the next person skips to a cache hit. That is the lazy cache filling. Two design choices I would defend in any review: “Ambiguous” is a first-class outcome. Two close candidates returns “did you mean the Seattle or the Dallas store” , not a coin flip. (This is the fix for Wall 4.) Resolution is fail-closed. The cascade runs and caches unfiltered, but nothing returns until it passes RBAC: can this user see the entity, its links, and any restricted properties. No scope and no trusted-caller flag means deny , not leak. Phase A: planning, stage by stage Dana’s question comes in. Phase A turns it into a plan without writing one character of SQL. User context assembly. Load who Dana is: role (sales rep), territory (Southwest), permissions (sales schema), saved vocabulary and preferences. Everything downstream is scoped by this. Intent classification (the one LLM call). Returns a structured verdict: intent type (a SQL query), entity types (CUSTOMER and the order entity), confidence, and a default join type. A second focused call pulls “Bike World” out as a value tied to CUSTOMER. A type gate then throws out any extracted value that does not fit the column’s type. Semantic resolution (the 2-pass vector search from Part 2). Turns “orders” into the order entity by meaning, not spelling , and flags ambiguity when two concepts score within a hair of each other. Scope resolution. Turns words like “my team” into row-filter hints, on pre-derived rules, in a few milliseconds. One subtlety I am careful about: this produces relevance hints, not security . Convenience and access control are different things, and conflating them is how you build something convenient and insecure. Routing (the deterministic core). Maps entity types to real tables, assigns each entity a role (the order entity is the subject, CUSTOMER is a filter) using a four-rule scoring system with no LLM , finds the foreign-key join path by traversing the domain graph, grounds “Bike World” to StoreID 620, and detects whether the query spans sources. RBAC Checkpoint 1. The first of two security gates. Coarse and early: drop any table the user may not see before the model ever sees the schema . Dana keeps her sales tables. Had she asked about HR salaries, that table would vanish here, and the model would never learn its column names. The output of Phase A is a frozen plan plus a warm entity cache. No SQL exists yet. Phase A decides everything that has to be decided safely, and hands the model a plan it would have to work hard to misuse. Two phase pipeline How to read it: the question flows top to bottom through Phase A (all blue, deterministic, except the one purple intent call), produces a frozen plan, and only then enters the Phase B loop where the model actually works. The two red boxes are the security gates, one per phase. Phase B: where the model actually works Phase B is a small state machine with three nodes: agent node: the LLM reasons and decides the next action tools node: executes exactly one tool, appends the result synthesize node: writes the final answer It loops between thinking and acting, then synthesizes. And it is bounded : at most 10 reasoning iterations and 15 tool calls per query, with a 30-second timeout per tool. Those limits are not decoration. They are the difference between “the agent retries a failed query once and succeeds” and “the agent burns your budget in an infinite loop of slightly-wrong SQL.” (This is part of the fix for Wall 7.) The agent has 11 tools but most queries touch two or three. A nice detail: resolve_entity looks like a live tool, but because Phase A already resolved Bike World into the warm cache, it usually just returns the cached answer if the confidence is above 0.5. The same operation is both "a tool the agent can call" and "something Phase A already did." Generating SQL without letting the model touch the dangerous parts This is where the two-phase philosophy pays off most concretely. Sql generate and execute How to read it: the model (the one purple box) writes only a slice of the query. Everything else is deterministic. The joins come from the graph, not the model. A join-tree builder traverses the foreign-key edges. No path between two tables means a loud error, never a guessed join. (This is the fix for Wall 2.) The model gets a whitelist of real columns and writes only the SELECT, WHERE, GROUP BY, and limit. Every column it returns is validated. A made-up or unqualified column rejects the whole generation. Unvalidated SQL never reaches the database. Filters get merged in by code, not the model. The load-bearing rule, again: the LLM never writes the joins. A made-up join is a silently wrong answer, and silently wrong is the worst failure an analytics system has. The fan-out trap (Wall 3, finally solved) Let me tell you about the bug that taught me to respect cardinality. Ask for total revenue per customer. The query joins customers to orders to line-items. The moment you SUM across a one-to-many join, the value repeats once per child row. A customer with one $100 order and five line items reports $500 . The SQL is perfect. The number is garbage. Nobody notices. So I built a deterministic planner that rewrites the SQL to prevent it. It keys entirely on the cardinality_class metadata from Part 2 (the PRESERVING vs MULTIPLYING tags), touches no table names, never calls the model, and fails open (if it cannot reason safely, it returns the original). Its strategies, picked by what the query actually references: Prune a multiplying join that is joined but never used (transitively, so chains collapse). Semi-join : a child used only in the WHERE becomes an EXISTS subquery, which filters without multiplying. Per-grain subqueries : two measures at two different grains get computed separately and joined back with a full outer join and null-safe matching. Count anchoring so a plain row count does not lie either. Why is this in deterministic code and not the prompt? Because “be careful about fan-out” is right most of the time, and “most of the time” is not a number you put on a financial dashboard. Cardinality correctness is provable from graph metadata, so I prove it. The second security gate (Wall 6) When the agent calls execute_sql, the SQL is correct but not yet safe. RBAC Checkpoint 2 runs here, and it is the fine-grained one. Where Checkpoint 1 decided which tables , Checkpoint 2 decides which rows and columns . It rewrites the SQL’s syntax tree: it injects mandatory and row-level filters straight into the WHERE. Dana's query silently gains a Southwest-territory constraint. A failed access check injects WHERE 1=0 , a silent deny that returns zero rows rather than leaking the existence of restricted data. Restricted columns are masked on the way out. The entire RBAC path is deterministic. There is no LLM anywhere in it , by design. Access control you cannot reproduce is access control you cannot defend. Then the agent synthesizes: it writes “Bike World placed 8 orders last quarter,” with any caveats (like “limited to your territory”) and a confidence, plus the SQL and rows underneath. The validation mesh Threaded through both phases is a layer I think of as a mesh, not a stage: eight validators, forty-four checks , firing at hook points all over the pipeline. Validation mesh How to read it: Three validators run in Phase A, and two of them (route, entity) can hard-abort the query before the agent loop ever starts. No point generating SQL for a query that was never going to work. The rest run in Phase B, guarding execution and the answer. Most checks are deterministic; a few are LLM calls (does the answer follow from the data), and those get a strict budget of three LLM validation calls per query . Two choices that matter: Every validator is wrapped in try/except. A validator that crashes, or is disabled, never blocks a query. Safety code that can take down the system it protects is a liability, so it is built to fail safe. Checks have severity tiers: block, warn (becomes a caveat, lowers confidence), or log-only. Not every problem should stop a query. Some just attach a footnote. Shipped vs. next (an honest naming note): the “memory validator” in this mesh has nothing to do with conversation memory. It checks whether a cached result went stale. The actual conversation is single-shot today . The read side of memory is wired (the system loads your saved vocabulary and scope on every query), but the write-back that would let Dana ask “and the quarter before that?” and have it understand “that” is built and not yet connected . Ask a follow-up today and it is treated as a brand-new question. I would rather say that than imply a multi-turn experience I have not finished. When the data lives in two databases at once (Wall 5) Some questions span sources: SQL Server sales joined to Postgres production. There is no cross-database JOIN here, and no copying one source into the other. A federated executor decomposes the query by source. It generates and runs a separate, independently RBAC-enforced query against each source, pulling only the rows that survive each source’s own rules. It joins the results in the application layer , in memory, by matching on entity keys. The requested join type (inner, left, right, full outer) is honored by deciding which side’s unmatched rows to keep. Each database only ever sees a query for its own data, scoped to what the user may see. The rows are stitched together only after each has been cleared. The data never leaves its home system. That property is non-negotiable. Why keep both a plan and a loop? Fair question: if Phase A plans everything, what is the loop for? They fail in different ways, and I wanted coverage on both. Job ( Deterministic plan (Phase A) : decide what : tables, entities, filters; ReAct loop (Phase B) : do it: write SQL, run, answer, recover) LLM? ( Deterministic plan (Phase A) : only intent classification; ReAct loop (Phase B) : yes: SQL text + prose) Gives you ( Deterministic plan (Phase A) : safety, reproducibility, auditability, speed; ReAct loop (Phase B) : execution, self-correction, tool choice) Costs you ( Deterministic plan (Phase A) : adaptability (the plan is frozen); ReAct loop (Phase B) : nondeterminism (must be fenced)) The loop earns its keep when the model generates SQL with a wrong column, the database errors, and the agent reads the error, regenerates, and succeeds. A straight pipeline cannot self-heal. The plan earns its keep by making sure the model can retry SQL all it wants but can never retry its way into a forbidden table or an invented join. The honest current limit: today the agent is locked into Phase A’s plan. It can fix a bad query, but not a bad plan. If routing picked the wrong table, retrying SQL will not help, because the agent cannot re-route mid-flight. The roadmap is to make the planning steps themselves into tools the agent can re-call inside the loop. That is the difference between a system shaped like an agent and one that is fully agentic, and it is the next real piece of work. Key takeaways Split the stochastic from the deterministic, and be ruthless about which is which. The model understands questions and writes creative SQL well. It is unaccountable at deciding access and inventing joins. Opposite sides of a hard line. Resolve the things a user names through a cheap-to-expensive cascade, and let “ambiguous” ask a question instead of guessing. A confident wrong answer is worse than an honest “which one?” Prove what you can prove with metadata instead of asking the model to be careful. Cardinality and access control are both provable from the graph, so they live in code, never in a prompt. Make your safety layer unable to take down the thing it protects. Wrap it, budget it, tier it. In Part 4 , I pick up the thread I keep deferring: documents. Everything here was structured data. Real questions often need a number and a paragraph from a PDF. The subject graph is meant to be the hub that ties the two together, and that is the most forward-looking, and most honest, story of the four. Next up, Part 4: Two Pipelines, Three Graphs, on fusing structured and unstructured retrieval with RAG. Never Let the LLM Write the Joins was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
Score: 28🌐 MovesJun 16, 2026https://pub.towardsai.net/never-let-the-llm-write-the-joins-fe5d1f1ccbe0?source=rss----98111c9905da---4 - How AI-Powered CMS Platforms Are Transforming Enterprise Content Operations
For years, enterprise content management was largely a publication tool. How do you get the right content, in the right format, to the right channel, without breaking workflows that span dozens of markets and hundreds of contributors? The answer was usually a combination of manual processes, siloed systems, and large coordination teams that grew historically […] The post How AI-Powered CMS Platforms Are Transforming Enterprise Content Operations appeared first on AI News .
- Apple @ Work Podcast: Leebry aims to unify your SaaS tools for AI use
Apple @ Work is exclusively brought to you by Mosyle , the only Apple Unified Platform. Mosyle is the only solution that integrates in a single professional-grade platform all the solutions necessary to seamlessly and automatically deploy, manage & protect Apple devices at work. Over 45,000 organizations trust Mosyle to make millions of Apple devices work-ready with no effort and at an affordable cost. Request your EXTENDED TRIAL today and understand why Mosyle is everything you need to work with Apple . In this episode of Apple @ Work, Dan Jaenicke, Director of B2B Product Strategy at MacPaw, joins the show to talk about Leebry .
Score: 28🌐 MovesJun 16, 2026https://9to5mac.com/2026/06/16/apple-work-podcast-leebry-aims-to-unify-your-saas-tools-for-ai-usage/ - Physical AI - Building Affordable & Sustainable Housing
The US faces a severe housing affordability crisis. Reframe Systems, founded by ex-Amazon engineers uses physical AI & robotics in micro-factories to solve this crisis.
Score: 28🌐 MovesJun 16, 2026https://www.forbes.com/sites/sabbirrangwala/2026/06/16/physical-aibuilding-affordable--sustainable-housing/ - AI can clone your voice. Why that's powerful—and dangerous
Using artificial intelligence, scammers can duplicate someone's voice with just seconds of audio, says the University of Cincinnati's Kimberly Hyun. Impostor scams are one of the most common forms of fraud, according to the Federal Trade Commission.
- Ex-Lush chief’s lawyers hike costs to ensure their AI model isn’t trained by juniors
Lawyers for the former chief executive of cosmetics giant Lush, Andrew Gerrie, have hiked their fees ahead of a trial this month, partly as they will not allow junior staff to train their AI model, City AM can reveal. A court heard this morning that Gerrie’s legal team from Brown Rudnick have pushed up a [...]
Score: 28🌐 MovesJun 16, 2026https://www.cityam.com/ex-lush-chiefs-lawyers-hike-costs-to-ensure-their-ai-model-isnt-trained-by-juniors/ - MiMo Code Is the Open Source Answer to Claude Code
MiMo Code Is the Open Source Answer to Claude Code DevOps.com
- Austin hotels struggle without convention center, even during SXSW. AI boosts smaller cities.
Austin hotels struggle without convention center, even during SXSW. AI boosts smaller cities. Austin American-Statesman
Score: 28🌐 MovesJun 16, 2026https://www.statesman.com/business/economy/article/austin-hotel-market-decline-texas-2026-22301274.php - This New Gadget Uses Custom Scents and AI to Improve Your Sleep
Kimba is an AI-powered sleep technology system that releases personalized scents to help you sleep better. Here’s how it works.
Score: 27🌐 MovesJun 16, 2026https://www.cnet.com/health/sleep/new-gadget-doesnt-just-track-sleep-uses-custom-scents-ai-improve-it/