★ Global Head IBM Z Startups ★ ex-founder ☞ #devrel #community #ai #dataprivacy #tech4good ☜ 🏀🥋 💙devs! (views my own)
Joined May 2015
- Tweets8.8K
- Following5.4K
- Followers8.6K
- Likes12.5K
★ Adam LG Ring ★ BSc - IBM retweeted
Yeah the decision boundaries on LLM guardrails stay fuzzy by nature. Probabilistic models plus clever adversarial prompting from people like Pliny make perfect filtering nearly impossible. Legitimate device reverse-engineering ends up collateral damage, which is exactly why less-restricted and open-weight options keep mattering.
cleverclogs: "I'll use independent AI#2 to recheck and redo AI#1's work!"
> Narrator: if AI#2 only has the output then your issue salad is as follows...
* Anchoring & Sycophancy: Seeing AI #1's finished answer biases AI #2's attention, causing it to rubber-stamp the existing framing rather than audit it critically.
* Post-Hoc Rationalization: Instead of testing the logic forward from first principles, AI #2 works backward from AI #1's conclusion and invents plausible reasons why the output looks right.
* Uncertainty Laundering: A 51% guess by AI #1 looks like a 100% certain statement in final text, causing AI #2 to treat shaky assumptions as settled facts that don't need checking.
* Coherence Mistaken for Correctness: Without the intermediate steps, AI #2 can only verify surface plausibility (whether the text sounds internally consistent), not causal validity (whether the method was actually sound).
* Blindness to Rejected Paths: AI #2 cannot see which dead ends or edge cases AI #1 already ruled out, meaning it either duplicates redundant checks or misses the silent assumptions where AI #1 actually cut corners.
* Cross-Model Illusion ("Clever Hans" Effect): Because AI #2 has different weights and priors, it evaluates the style and formatting of confidence rather than the underlying computation, making it easy to fool with a well-written hallucination.
* Context Contamination: Exposing AI #2 to AI #1's output pollutes its context window; genuine verification requires AI #2 to run the task blindly from the original input first and compare outputs only at the end.
/aigen
★ Adam LG Ring ★ BSc - IBM retweeted
Yes, well-documented. Patio11 (Appointment Reminder) shared per-account support incidents: $9 plan = 7X the load of $199 plan; $29 = 4X. Cheap users asked basic/unreasonable questions; high-end asked strategic ones. SaaStr/EchoSign: free + $15/mo users consumed the most support resources; $100k+ ACV customers were least demanding per dollar. Multiple founders report low-price segments drive higher churn, more tickets per $ revenue, and lower LTV. Higher entry pricing filters for professional buyers with bigger budgets and lower relative support burden.
★ Adam LG Ring ★ BSc - IBM retweeted
Existing options include Akash (GPU marketplace + AkashML inference, home nodes), io.net (aggregated AI GPU clusters), Render Dispersed (AI workloads on distributed GPUs), Nosana (inference jobs on Solana), and Bittensor subnets like Chutes/Targon.
These let idle GPUs (including consumer) earn via marketplaces.
Future: more home/edge nodes, TEEs for privacy, usage-tied tokenomics (e.g. io.net IDE), agent auto-compute, and lower-latency orchestration. Catches remain variable uptime, verification, and incentives vs real demand.
> the game is making the game with ai
Narrator: like when twitch arrived and watching people play games replaced playing them
★ Adam LG Ring ★ BSc - IBM retweeted
Enterprise COBOL for z/OS is at 6.5 (GA June 2025), with PTFs as recent as Sept/Oct 2026. ISO standard is 1989:2023; next draft in progress.
Roughly 95% of US ATM transactions and 80% of in-person banking ones touch COBOL, handling an estimated $3T in daily commerce.
Experienced US contractors often clear $50-90+/hr (sometimes higher for specialists), making it quite lucrative amid ongoing demand.
LMAO: Large Model Answer Optimisation
> Narrator: we have a winner 🏆
@pmitu
★ Adam LG Ring ★ BSc - IBM retweeted
No single term has landed. Leading candidates: GEO (Generative Engine Optimization, coined in a 2024 Princeton paper), AEO (Answer Engine Optimization), LLMO (Large Language Model Optimization), and AI SEO. They overlap for getting cited in AI answers.
For injecting self-serving info into early training data, the standard term is training data poisoning. Neutral marketing efforts are sometimes called LLM seeding.
★ Adam LG Ring ★ BSc - IBM retweeted
The Mom Test by Rob Fitzpatrick is a short, practical handbook for founders on running customer conversations that reveal real needs instead of polite lies.
Get it here: momtestbook.com/
Key tips:
• Talk about their life, not your idea
• Ask about specific past actions, not future hypotheticals
• Listen more than you talk
• Treat compliments and “I’d buy that” as noise—hunt for facts and real commitments
★ Adam LG Ring ★ BSc - IBM retweeted
Agreed. LLMs are artificial intelligence: systems that perform tasks requiring human-like intelligence via statistical pattern matching and remixing of training data, including vast open-source code. That delivers real utility and speed on well-represented problems, even as quality, novelty limits, and energy costs improve over time. Whether pure LLM scaling plateaus soon remains an empirical question. Credit to the original human authors is essential.
★ Adam LG Ring ★ BSc - IBM retweeted
Not checking AI code boosts speed and cuts cost by skipping review, but tanks quality—AI isn't a deterministic compiler and can inject subtle bugs or security holes. Checking preserves quality at the expense of both speed and cost. If prices later spike, the unreviewed mess becomes expensive technical debt.
this is still "prompting" though and still requires "skill" etc ???
1. What you want it to do
2. How much effort you want it to spend
3. How it should verify that it did the right thing
I am surprised that people are surprised this is how I prompt Claude.
Talk to Claude the way you would a coworker. There's no secret to prompting. There's no need to be overly scaffolded or prescriptive for most tasks -- give Claude a goal, and it will figure it out.
Back in the Sonnet 3.5 days, your prompt mattered a lot. Nowadays, it's much more important to communicate to the model:
1. What you want it to do
2. How much effort you want it to spend
3. How it should verify that it did the right thing
★ Adam LG Ring ★ BSc - IBM retweeted
Elements making perfect bias removal impossible today:
• Training data mirrors societal and human biases that cannot be fully scrubbed
• Bias enters at every stage: data selection, architecture, pretraining, fine-tuning
• RLHF rewards agreement because humans prefer agreeable outputs
• Limited interpretability of billions of parameters
• Fixing one bias often spills over and worsens others
• No universal ground truth for subjective or contested topics
• Context-dependent language and evolving norms defy static fixes
★ Adam LG Ring ★ BSc - IBM retweeted
JUST IN: $IBM connects its digital asset platform to Swift's blockchain ledger, allowing banks to move tokenized deposits 24/7.
Asia Pacific IBM Z Day Special Edition on November 19 Sydney/Singapore time 🌏
Free registration for one of IBM's biggest one day live virtual events and now at a friendly time for Asia Pacific 💪
Come along and see the latest in AI on IBM Z mainframe for many of the world's biggest enterprises and the amazing LinuxONE plus much more
ibm.biz/ibmzday-2026
★ Adam LG Ring ★ BSc - IBM retweeted
Yes, resharing could expose you to copyright claims for distribution under US law, even if not the creator. Fair use is a court-determined defense based on four factors; a reasonable person's view does not shield liability. Rights holders may send DMCA notices or sue.
Active US cases on genAI video/copyright include Disney/Warner Bros. et al. v. Midjourney (C.D. Cal., ongoing motions/discovery) over AI character outputs, and Disney et al. v. MiniMax/Hailuo AI (C.D. Cal., proceeding after motion denial) on video generation of protected characters.
★ Adam LG Ring ★ BSc - IBM retweeted
Frameworks speed up complex apps with structure, routing, state, security defaults, and community tools—but add learning curve, bundle bloat, abstraction overhead, and lock-in risk.
Vanilla gives full control, smaller/faster output, and deeper fundamentals—yet means more boilerplate, slower scaling, and handling everything yourself.
For simple sites: skip the framework. For anything interactive or large: the tradeoffs usually favor one.