Quintin G.

Big news in the AI world. A federal judge just ruled that Anthropic’s use of legally purchased books to train its Claude model qualifies as fair use. This is the first major court decision to say so. This starts to draw a real line in the sand: how you acquire your training data matters. As someone building AI in a highly regulated space (healthcare + legal), rulings like this are huge. They help set clearer expectations for how to build responsibly, and make smarter choices about data strategy from day one. Thoughts? https://lnkd.in/ecTne6GP

19

1 Comment

Boundless Opportunity in Florida (3rd largest state) for Construction Equipment OEMs, Dealers and Rental Companies Based on BiltData.ai's analysis of 100M+ transactions, BiltData has examined the diverging paths inside the U.S. construction equipment and rental industries, and shows OEMs, dealers and rental companies where opportunity resides. Florida is powerhouse, the 3rd largest among U.S. states in total construction spending by 2030, trailing only Texas and California. The state currently represents ~7% of all USA construction spending in 2025, and is projected to grow to 8% by 2030 driven by Florida's rapid population growth, infrastructure needs, and data center projections. By example, Florida is poised for substantial data center growth between 2025 and 2030, transitioning from a secondary market to one of the fastest growing U.S. hubs. This surge is fueled by AI hyperscale demand, Florida’s sales tax exemptions on data center equipment and electricity (for facilities with >100 MW). Most importantly, utilities are investing billions in hardening and new power generation. Florida's Top 5 MSAs (metropolitan statistical areas) in 2030 account for ~80% of the state's construction spending: 1. Miami-Fort Lauderdale-West Palm Beach 2. Tampa-St. Petersburg-Clearwater 3. Orlando-Kissimmee-Sanford 4. Jacksonville 5. North Port-Bradenton-Sarasota Florida's powerhouse economy directly boosts demand for construction equipment, particularly earthmoving. Biltdata’s predictive buyer intelligence has identified significant construction equipment sales, rental and product support opportunities identifying a small percentage of contractors who will will drive ~45% of demand. BiltData.ai's predictive buyer ops turns raw data into actionable intel for construction equipment OEMs, dealers, and rental companies, optimizing marketing, inventory, and sales for growth in high-demand sectors. Our proprietary BiltReady Opportunities provides intel with a 100% view of actionable opportunities: Sales, Rental, Parts & Service. Behind BiltData is a proprietary AI engine trained on 100M+ transactions. No Opportunity. Left Behind.

14

The scale of uncovered fraud in reports like this isn’t surprising to anyone inside government — because the root cause is structural, not behavioral.   Agencies are being asked to fight cross‑program, cross‑jurisdiction fraud with visibility that stops at their system boundary. The signals exist, but they’re trapped in eligibility systems, payment platforms, identity stores, and case management tools that were never designed to correlate with each other.   Fraud isn’t “hidden.” It’s uncorrelated.   Until agencies can see operational, identity, and behavioral signals in one place, fraudsters will continue to exploit the seams between programs. This is why modernization efforts stall, why improper payments persist, and why oversight bodies keep finding the same patterns year after year.   The mission need is clear: unified operational intelligence across systems, not more point solutions inside them. #ProgramIntegrity #FraudPrevention #ImproperPayments #FederalModernization #DataCorrelation #OperationalVisibility #InteragencyCollaboration #DigitalGovernment #EnterpriseArchitecture #MissionSupport https://lnkd.in/eCYRsirV

3

Yesterday I wrote about authority over AI influenced decisions inside institutions. This week we are seeing the sovereign version of that question. Reports describe a standoff between the U.S. Department of Defense and Anthropic regarding model usage constraints for military applications. Under the Defense Production Act, the U.S. government holds broad authority to compel private firms in matters of national security. This is not solely an AI safety story. It is also a governance signal. As AI systems become embedded across defense, energy, telecom, healthcare, finance, and logistics, they are no longer just software tools. They are becoming intelligence infrastructure. And infrastructure, when it intersects with national security, sits under sovereign authority. Boards are not modeling this adequately. Sovereign override risk is the risk that a state can compel, reshape, or suspend corporate AI governance frameworks under emergency or national security powers. This risk exists independently of technical performance, internal policy, or contractual intent. Deployment guardrails can become conditional when sovereign authority is invoked. For boards and executive teams, the questions are structural: What percentage of our AI capability depends on vendors subject to national security compulsion? In which jurisdictions do those vendors operate, and under what emergency authorities? How would our governance posture change if model policies were altered by sovereign mandate? AI governance is no longer purely technical or ethical. It is geopolitical. We have already seen sovereign influence reshape capital markets through export controls, sanctions, and emergency authorities. That structural influence is now surfacing in the intelligence layer. When AI becomes infrastructure, it becomes subject to sovereign power. Boards that treat AI risk as limited to bias, drift, or safety are modeling only one layer of exposure. The structural layer, sovereign authority over intelligence infrastructure, is now clearly visible. Visible risks should be measured.

19

1 Comment

It's hard to cover the fallout between the US Department of Defense and Anthropic without diving dangerously close to politics, but here goes. For those that missed it: -- At the heart of the standoff was a technical "red line." Anthropic refused to remove safeguards against using Claude for fully autonomous weapons, amongst other things, arguing that frontier models are "simply not reliable enough" to remove humans from the loop without risking lives. -- Unlike everything that has come before, AI doesn't follow a repeatable workflow. In a low-stakes environment, that might mean a weirdly worded email. In a high-stakes environment (like the ones the DoD is navigating), that unpredictability can lead to catastrophic failure. Law or no law, red lines or not, anyone working with these models knows that AI is currently an *incredibly powerful* assistant, but even in agent form, not a totally autonomous autopilot. But as I’ve said before: Trust is not a strategy. Whether it’s a battlefield or a sensitive corporate database, we have to move away from a default posture of "granting access and see" to one of lowest necessary access and/or human-in-loop as the risk profile of the assignment grows. Public service announcement part over, what I don't understand is the "supply chain risk" designation from the US Govt, or how OpenAI stepping in with the same set of red lines is any different. We are living in a world where ~£3000 gets you usable, local inference. Paired with open models that are quickly closing the capability gap to frontier models. So are these "red lines" from model vendors a real protection or just clever marketing? What do you think is really at play here? And don't forget: If you are granting an agent access to your primary systems without restricting certain data or keeping a human-in-the-loop, you aren't just testing technology. You may be outsourcing your judgment. #AI #CyberSecurity #DefenseTech #Anthropic #GenerativeAI #DataPrivacy

8

3 Comments

There's approximately seven types of ASR failure modes we see every week across voice agents. 1- Noise-Driven Omissions: Background noise drops key entities (dates, numbers, names etc) 2- Substituted Intents: “Cancel” becomes “schedule” (this normally happens when the signal quality drops or background inference) 3- Formatting Drift: The transcript is correct, but the format silently changes (120 → one-two-zero) 4- Truncation / Endpointing Errors: The ASR cuts callers off mid-sentence 5- Hallucinations: ASR inserts phrases the caller never said 6- Accent & Dialect Variability: This is caused by training and testing voice agents with non-diverse recognition data 7- Silent Regressions: Vendor/model updates shift behavior with no visible code changes What ASR failures have you noticed? Let's debug :)

25

4 Comments

National security customers are focused on transforming their ISR strategies. Our team did some digging and we realised that there was a global trend across countries who were all looking to change the way they leveraged space based GEOINT. There were several common themes that emerged; - Organisations are moving toward procuring data from multiple commercial sources, - Challenges persist around effectively co-ordinating sovereign assets with commercial capture capabilities, - Harmonizing satellite datasets from multiple sources at scale is an issue, - Securely sharing large quantities of spatial intelligence data across teams and with allies was a pain point. The global intelligence community is starting to invest in Earth Observation Data Infrastructure (EODI) to scale their operations in order to produce better intelligence products for people that need them most. #EODI #geoint #nationalsecurity #space

20

In Mackenna's Gold there is a scene where Omar Sharif's character Colorado points a gun at Gregory Peck's character Mackenna asking where the map to "cañon del oro" is. Peck nonchalantly replies "I burnt it". And repeats it. Then Sharif points his gun to Peck's head and says "you might have burnt the map but I know it is here" (i.e. in Peck's head; he has memorised it). I was thinking of this scene in the context of what Intellectual Property means in the age of AI and vibe coding. With code becoming relatively throwaway / easy to write, a company's IP is increasingly what is in the founders' / employees' heads! It's the know-how of what to build and how, and not necessarily the precise instantiation of this build. This possibly explains the new trend of acquisitions of AI companies, where acquiring companies simply hire one or more founders of "acquiree" companies, and the rest of the company goes its own way.

25

1 Comment