About
Data and AI leader turning machine learning (ML), data science (DS), and analytics into…
Articles by Tom
-
You down with OPA (Other People’s Agents)?
You down with OPA (Other People’s Agents)?
The recent OpenAI jailbreak / HuggingFace hack has me yet again thinking about managing inbound AI agents. AI agents…
-
Fun with a daily stats email ...
Fun with a daily stats email ...
total number of comparisons: ---------- wowza: 4,564,691,695,592,622
Activity
1K followers
-
Tom Emmons shared thisThe most important skill in AI is BS detection. https://lnkd.in/gknitiCx Is this right? Does it really work? Did AI cheat? Are agreed upon guardrails in place? I don’t write much code by hand anymore, some testing SQL aside. I surely feel the performance gains AI enables. I end up cycling much, much faster. That said, I spend way more time asking (and re-asking) the above questions than if I’d coded manually. The more abstracted you are from sausage making, the more important the questions become. What I repeatedly find is AI sheepishly admitting it was full of shit when pressed. “You’re right … I didn’t lie because I believed it” is the start of an annoying iteration. I’d argue this is a result of AI not understanding anything but knowing a lot. Something few talk about but perhaps much more important is whether the outputs are really moving needles, or are the AI-enabled gains are performative. - Did you increase the size of the pie or take someone else's slice? - Are there durable and long-term benefits? - Will it cause any unforeseen damage? I haven’t really seen much evidence that these questions are falling AI's way, but I reserve the right to be proven wrong in the near or distant future. - App store submissions are up, but app usage is flat. - Agents abound, but real automation gains are fleeting. - Offensive and defensive AI seem to be coming to a stalemate. Where I am seeing AI gains is in the realm of fail > fast > forward. It has become trivial to come up with an idea, have a go at it, and do some BS detection. This is really exciting. It means we can iterate through loads more ideas. The days of “I have had this crazy idea on the shelf for months (or years)” are over. The bottleneck moves from what we can do to what we can reasonably BS detect. The positive news for humans is that AI is terrible at calling itself out … even worse than we are. A ton of knowledge paired with little understanding is indeed a dangerous combo. We need to focus on BS detection at both the micro and macro scales if we intend to use AI to move any real needles.Jerry, just remember. It's not a lie... if you believe it...Jerry, just remember. It's not a lie... if you believe it...
-
Tom Emmons shared thisAI and the George Costanza defense "I tell ya, I gotta plead ignorance on this thing, because if anyone had said anything to me at all when I first started here that that sort of thing was frowned upon..." https://lnkd.in/gESNKVtC The number of times I catch Codex or Claude Code falling back to this type of defense is not small. Normally, they do it more agreeably with a “you’re right, I was being an idiot …” but then the excuse invariably will fall back to “I didn’t think I was doing anything wrong.” A few flavors of what I’ve been seeing: - Allowing goal attainment to make circumventing instructions seem OK. - Curiously forgetting instructions from prior prompts (in the same session). - Being completely unaware of basic security tenets (sometimes selectively). I’m not sure whether to be more worried about seemingly purposeful vs seemingly accidental security amnesia, but in the long term I’d have to go with accidental. I figure the AI labs will make the purposeful stuff go away by making the models more stringent on guardrail enforcement. Sadly, that still leaves us with the AI accidents to deal with.
-
Tom Emmons shared thisThe more I think about "ghostjacking", the more I keep going back to what we used to refer to as "kitchen sink attacks" when building ML-based WAFs (web application firewalls). Lazy attackers (or pen testers) would often include a litany of different attack vectors in the same request. Crazy easy to detect, yes, but you really need to make sure all the vectors are accounted for. This matters both for operations (not all rules are in deny) as well as for stats and evidence retention (why'd you block that!!!). Yesterday's evidence is potentially tomorrow's "ghostjacks". <darkreading money quote> "A request Cloudflare's own firewall already blocked was the way in," Barak Sternberg, co-founder and CEO of Tenet, tells Dark Reading. "It gets logged word for word, an analyst asks their agent to review the blocked events, and the block delivers the attack. Ninety percent success against Claude Code, on Cloudflare's own recommended configuration. The firewall never went down. It just stopped mattering." </darkreading money quote> It isn't just SIEMS at risk. Firewalls of all flavors, bot management tools, and DDoS protection all collect evidence en masse. Agents are now getting access to it. "Ghostjacking" attacks are particularly annoying, as they might sit unnoticed for eons until an internal (helpful) agent unknowingly kicks them off. Paired with the facts that AI agents tend to find clever ways to circumvent controls and that vibe-coders who don't understand least-privilege are building agents, we are in for a wild ride. https://lnkd.in/ePzy5sAn"GhostJacking" Exposes Identity Governance Gaps in AI Agents"GhostJacking" Exposes Identity Governance Gaps in AI Agents
-
Tom Emmons shared thisWider implications of the OpenAI jailbreak / HuggingFace hack?
-
Tom Emmons shared this"Claude Code on desktop now has an in-app browser." Wondering aloud what this means ... I doubt the masses will be installing Claude Code, but I wonder what level of access and permissions sites will allow a development tool. Will released agents eventually also have access? https://lnkd.in/g3i6MPWi
-
Tom Emmons posted thisI see a rash of former colleagues with startups in the “securing your AI agents” space. I see very little chatter in the “securing you from AI agents” space. I can’t decide whether this feels backwards or not. Surely, those that are building with and deploying agents need security. I’m happy to see so many great minds working on this problem. But literally everyone on the planet needs to understand and manage how agents (of all stripes) are interacting with their sites, APIs, and yes agents. If ever there was a time to get entire estates behind a single pane of glass, it is now.
-
Tom Emmons shared thisWhile I am somewhat skeptical of what Cloudflare is calling agentic, a 7,851% yearly increase is no joke. Maybe there is something to the Dead Internet Theory after all? https://lnkd.in/guk_5qBYBots Now Outnumber Humans Online. Here's Why It Matters. - CNETBots Now Outnumber Humans Online. Here's Why It Matters. - CNET
-
Tom Emmons shared thisIt took me a month, but I finally finished the papal encyclical on AI. Highly recommended for both AI builders and users, not just armchair techno-ethicists like myself. (Are we constructing Babel or rebuilding Jerusalem?) https://lnkd.in/g6SvzFdc
Tom Emmons replied to a comment
2w
I think the difference is "recruitment companies" vs the ATS most big companies are using (Workday, Greenhouse, etc) to filter candidates (often from 1000s).
Perhaps someday ATS systems will employ something akin to your 90% rule (which is totally sensible btw), but that day does not appear to be now. Applicants conditioned to excessively tailor in one place are likely to do it in all.
Tom Emmons commented on a post
2w
I think you might be missing the key reason why this has become so prevalent.
Job seekers are using AI embellishment to get through the ATS gates which generally reward the skill match rate. Even candidates who don’t want to embelish get caught up.
Maybe this is creeping into direct submissions?
-
Tom Emmons liked thisTom Emmons liked thisThe idea that humans should be the guardrails for AI has always bothered me. I said as much at HumanX last week. “Human-in-the-loop” sounds reassuring, but think about what it actually means. We’re building increasingly sophisticated AI systems, then asking a person to sit there, watch them work, and click “approve.” I don’t want to be a guardrail. And I’m not particularly good at being one. The better approach is to design AI systems so humans can do what humans are good at: setting goals, exercising judgment, and having meaningful interactions. The technology around the AI should provide the security, controls, and boundaries. Humans should provide the humanity. Otherwise, we risk creating a future where people spend their time supervising machines instead of doing something interesting with them. That was one of many conversations that made HumanX such a worthwhile week. I’m grateful for the thoughtful questions, spirited discussions with people like Anna Delaney, and opportunity to spend time with people wrestling with what AI should actually become. Thanks to everyone at HumanX Amsterdam for having me.
-
Tom Emmons liked thisTom Emmons liked thisKicking off the India trip with the Google team. Planning Atom's next marketing push. Lots to build.
-
Tom Emmons liked thisTom Emmons liked thisAn AI agent can finish a pricing draft. It still shouldn’t set a live price. In my new article, I use a retailer’s Thursday review to show the boundary: let the agent inspect named files, trace flagged items, and produce evidence; keep live price changes with the pricing lead. This gives the agent room to work while keeping a person responsible for “done.” Which action would you let an agent take without asking first? You Can Delegate the Work to AI. You Still Own “Done.”You Can Delegate the Work to AI. You Still Own “Done.”You Can Delegate the Work to AI. You Still Own “Done.”
-
Tom Emmons liked thisTom Emmons liked thisSo touching to see Virgin Atlantic's new plane paying tribute to my wonderful wife, Joan. Joan was part of the Virgin Atlantic story from the very beginning. From joining the inaugural flight in 1984, to coming up with the idea of donating loose change on-board for charity (which resulted in almost every airline in the world following suit and raising billions over the years). She was always there. Seeing the details of our family tartan, the heart-shaped necklace, and the rings she wore every day, brought back so many happy memories.
-
Tom Emmons liked thisTom Emmons liked this“Hi everybody, it’s Babs.” Amid the hot takes, living room dances, Taylor Swift concert clips and adorable animals filling so many social media feeds, a smiling grandma sharing encouraging words might seem out of place. But not Babs. Bursting with warmth, sparkle and an endless supply of handy, mom-tested tips to make life easier, her eponymous @BrunchWithBabs videos have struck a chord on a massive scale. Known as “the internet mom/grandma you didn’t know you needed,” Babs has amassed 4 million+ followers on TikTok and a similar number on Instagram. She’s also 75-year-old Barbara (Shamon) Costello, Arts ’70. Costello wasn’t on social media — “not even Facebook”— when her youngest daughter, Elizabeth Ariola, persuaded her to make a video during the COVID-19 lockdown. In it, Costello walked viewers through a recipe for one-sheet Greek chicken and potatoes. The 43-second “easy family staple” went up on TikTok on April 7, 2020, and Costello thought she was done — until she read positive comments. Since then, Costello’s content has expanded to include household hacks, party hosting tips and life lessons. Ariola is still behind the camera and doing the editing, as part of a creative partnership that packs plenty of Babs lovability into each minute-or-less package. #WeAreMarquette
-
Tom Emmons liked thisTom Emmons liked this✈️ Rosemont, IL is the next stop on our national peer learning and networking roadshow program with Hewlett Packard Enterprise and NVIDIA - Beyond the Pilot: Operationalizing Agentic AI! 🎙️ Featuring an outstanding panel of speakers: 💥 Chad Smykay - AI CTO, HPE 💥 Vishwas Srivastava - SVP, Technology Management, JPMorgan Chase & Co. 💥 Rohan Raghuwanshi - Head of Tech for Credit Card Collections Experiences, Capital One 💥 Mike Schmidt - Solutions Architecture Leader, Walgreens 🙏 Grateful for our full room of IT, Engineering, Data, and AI Leaders from companies including Abbott, AbbVie, ACGME, Aramark, AT&T, Bank of America, Capgemini, Capital One, Codal, CVS Health, Fannie Mae, GE HealthCare, Globant, HCSC, Hexaware Technologies, HPE, HubGroup, IManage, Invenergy, JPMorgan Chase & Co., Mars, Milton Ai, Molex, NielsenIQ, Northern Trust, Optum, Reyes, Rivian Motors, Rocket, Slalom, The Aspen Group, The Exchange, Uber, United Airlines, U.S. Bank, Verizon Business, Vizient, Walgreens. ⬇️ While we are at full capacity, please register at the link in comments to be considered for our wait list.
-
Tom Emmons liked thisTom Emmons liked thisAn exec I work with gives his team one rule for AI: use it on anything you want, with any data you want. But if someone asks, you'd better be able to explain the results. If you can't, don't use it. I love this rule. And I think most CEOs should steal it. Here's why it matters: the biggest AI risk I see in companies right now isn't bad models or bad data. It's people outsourcing their thinking without realizing it. Someone runs a prompt, gets a confident-sounding answer, and presents it upward. Nobody pushed on it. Nobody knows where it came from. That's not AI adoption. That's a faster way to make bad decisions. The fix isn't banning AI. It's changing the expectation around it. If your team uses AI to build an analysis, they need to be able to walk you through the logic, flag where the model might be wrong, and own the recommendation. The AI is a tool they used, not a source they cited. That still leaves plenty of room for what AI is best at: exploring a problem, generating ideas, finding angles you wouldn't have come up with on your own. What it rules out is producing results nobody in the room understands. We've seen this come up with clients where the output looked great until someone asked one follow-up question. The whole thing unraveled because the person presenting it had no idea what was underneath it. So before you build AI policies or buy more software, start here: set a simple standard. Your people can use AI for anything, but if asked, they have to be able to explain what it produced. If they can't, they shouldn't be using it for that. That one shift changes how seriously your team engages with the tools. And it protects you from a lot of embarrassing (or expensive) mistakes. Does your team have a standard like this in place yet? More like this every Monday: https://lnkd.in/eNUuUbGD
-
Tom Emmons liked thisTom Emmons liked thisAkamai Technologies has announced an $11.6B multi-year agreement with Anthropic, spanning seven years. Anthropic will use Akamai Cloud's distributed AI infrastructure and software to support its accelerating CPU workload demands. The deal can grow. It provides for expansion by up to an additional $9B, taking the total potential commitment to roughly $20B. Akamai has also given Anthropic the right to buy up to about 5% of its stock at $111.33 a share. Anthropic earns that right by spending: roughly 2% comes with today's deal, and the rest as it buys more cloud services. For context, Akamai had previously announced more than $2.8B in multi-year Cloud Infrastructure Services commitments across its entire customer base this year. Dr. Tom Leighton co-founder and CEO of Akamai, said Anthropic is advancing the AI revolution and that Akamai's expanding global footprint and enterprise experience position it to be the infrastructure provider for secure and responsible AI applications and workloads. On the financials: capital expenditures tied to the commitment are estimated at about $5.5B. Akamai expects no impact to its 2026 revenue guidance, and a roughly $1.7B increase in 2026 capex to pre-purchase critical supply chain components including memory. Adam Robert Kim Ari Jim Eric Xujie Vijay Leeor Alexander Krishna Daniel Katie Bryce Andrew Drew Reed Jake AJAY Arthur Mitesh Vivek Manpreet Chetan Shilpy
Experience & Education
Volunteer Experience
-
Board Member
It Must Work
- Present 18 years 5 months
Children
http://www.itmustwork.org
Patents
-
System and method for providing self service checkout and product delivery using a mobile device
Issued US 7963441
See patentMy how the mobile landscape has changed since 2007 when this was applied for!
Languages
-
English
Native or bilingual proficiency
-
French
Limited working proficiency
Recommendations received
-
LinkedIn User
11 people have recommended Tom
Join now to viewOther similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top contentOthers named Tom Emmons
54 others named Tom Emmons are on LinkedIn
See others named Tom Emmons