Dr. T. Justin Webb

U. S. Army Officer, Ph.D. in Psychology, Author

[Dr. T. Justin Webb @webb8020]

Distinction: consolidation can cut marketplace sprawl and still thin oversight. Accept or reject: funding moves without a prove-the-claim test are incomplete HITL. Reply ACCEPT or REJECT plus one reason.

[Dr. T. Justin Webb @webb8020]

Fair read: consolidation cuts marketplace sprawl, not Authority of Correction. Org charts move funding; they do not replace prove-the-claim. After DRPM-UxS stands up, what single test shows oversight got faster without getting thinner?
https://t.co/paPg9qKstc

[Dr. T. Justin Webb @webb8020]

@DefenseScoop Forced choice for the thread: if the rewrite keeps only one non-negotiable tonight, which is it: termination, geographic delimitation, or civilian-harm analysis in DAWG? Pick one and say why in a line.

[Dr. T. Justin Webb @webb8020]

@DefenseScoop Fair read: 90-day 3000.09 speed is not the scandal; hollowed safeguards would be. Distinction: review cadence is not safeguard content. Which stays non-negotiable in the rewrite: termination, geographic delimitation, or civilian-harm analysis in DAWG?

[Dr. T. Justin Webb @webb8020]

@BreakingDefense @AccentureFed Fair read: open models + swarm tempo break binary HITL; risk-scale autonomy fits. Distinction: scale is not Authority of Correction. Kinetic needs rejectable evidence, not a polished brief. What artifact rides with each agent COA before refuse of a fluent wrong option?

[Dr. T. Justin Webb @webb8020]

Three models cleared 79%+ on a public production-parity eval. Then they failed an eliminatory hallucination gate. A composite score is not an acceptance criterion.

Here is the incident, not a product ranking.

One fact: a public Hugging Face study replayed production-parity conversational-agent sessions across eight models. Three models scored above 79 percent Final Score, then were rejected for exceeding a Hallucination Rate threshold treated as an eliminatory gate independent of the weighted average.

Source (scout-cited):
https://t.co/B97taKGLYx

One distinction (AI literacy, not a leaderboard dunk): what AI is, in this case, is a system that can look ready on a composite slide while failing a safety property that matters when agents decide on real work. Vendor claims love averages. HITL starts with deciding which properties are hard fails before anyone buys the briefing slide. Impact levels are the difference between a score and an acceptance gate.

One question (kill question): Is hallucination (or any safety property) a hard fail in our acceptance criteria, or can a high composite score paper over it?

Winner anatomy: score the mission traits, keep safety as a separate filter, ask the kill question before the slide becomes policy.

Paid corpus license: search free; agent 0.05 USD/page; training 5000 USD/year.
Book: AI-Powered Operational Design, 2nd Ed -- https://t.co/qMmB9EplTQ ($43.96).
Also: AI-Powered Strategic Thinker 3 -- https://t.co/zRsjEPonXE ($19.99 paperback).
Follow @webb8020. Link in bio: Pensator Strategus | AI for GIs | FREE.

[Dr. T. Justin Webb @webb8020]

#1 on LMArena. #4 on academic benchmarks. Same model. Both rankings accurate. Procurement memo said "industry-leading." Which leaderboard? Which methodology? Every scoreboard has different criteria. #MilitaryAI #AIforGIs

Free course at Pensator Strategus. Link in bio.

[Dr. T. Justin Webb @webb8020]

50% fewer hallucinations. Staff stopped checking. Battalion reported full readiness at 70% strength. Commander cited it in a decision brief. "Reduced" without a baseline hides the residual risk. #MilitaryAI #AIforGIs

Free course at Pensator Strategus. Link in bio.

[Dr. T. Justin Webb @webb8020]

U.S. banned advanced GPUs to China. No hardware, no frontier AI. That was the assumption. DeepSeek achieved frontier performance anyway. 10-20x cheaper. Architectural innovation routed around hardware restrictions.

Free course. Link in bio. #MilitaryAI #AIforGIs

[Dr. T. Justin Webb @webb8020]

128K to 10 million tokens. 78x expansion. Largest generational context jump ever. A planning team went from segmenting intelligence to processing 15,000 pages in one query. Patterns surfaced that were buried for weeks. #MilitaryAI #AIforGIs

[Dr. T. Justin Webb @webb8020]

115,000-token OPORD. 128,000-token window. 13K headroom. System crashed anyway. Hidden overhead from system prompts ate the buffer. Three annexes silently truncated. Add 20-30% buffer to every calculation. #MilitaryAI #AIforGIs

[Dr. T. Justin Webb @webb8020]

Cheapest AI platform. Three drafts. Two hours rewriting by hand. Her peer used premium. First draft. Done in 18 minutes. Cost-per-token is not cost-per-outcome. Staff time erased every penny saved. #MilitaryAI #AIforGIs

Free course at Pensator Strategus. Link in bio.

[Dr. T. Justin Webb @webb8020]

GPT-5.2 for daily SITREP summaries. A budget model produces identical output at 1/28th the cost. 95% of token volume did not need premium reasoning. Single-platform standardization is not optimization. #MilitaryAI #AIforGIs

[Dr. T. Justin Webb @webb8020]

Signed a 12-month AI contract after thorough evaluation. Six months later 3 competitors released faster models at half the price. No quarterly review. No exit clause. Platform selection is not a one-time decision. #MilitaryAI #AIforGIs

[Dr. T. Justin Webb @webb8020]

G6 said pick one AI platform. G2 picked the familiar one. Three weeks later drone footage sat unanalyzed. Text-only AI on a multimodal mission. Match the platform to the mission. #MilitaryAI #AIforGIs
Free course. Link in bio.

[Dr. T. Justin Webb @webb8020]

One staff section ran 5 AI platforms. Peers called it chaos. Then the quarterly review showed who delivered better intelligence at lower cost. Standardization is not optimization. #MilitaryAI #AIforGIs

[Dr. T. Justin Webb @webb8020]

109B parameters. Only 17B fire per token. Llama 4 Scout fits on one H100 GPU. Frontier capability on modest hardware. But your team owns every security patch, every update, every 0200 failure. #MilitaryAI #AIforGIs
Link in bio.

[Dr. T. Justin Webb @webb8020]

50% fewer hallucinations. Staff stopped checking. Battalion reported full readiness at 70% strength. Commander cited it in a decision brief. "Reduced" without a baseline hides the residual risk. #MilitaryAI #AIforGIs

[Dr. T. Justin Webb @webb8020]

Cheaper AI model. Three attempts. 55 minutes. Premium model with visible reasoning. One attempt. 12 minutes. The cheaper platform cost less per token but more per task. Iteration overhead is the expense nobody tracks. #MilitaryAI #AIforGIs

Pensator Strategus. Link in bio.

[Dr. T. Justin Webb @webb8020]

80.9% code accuracy. Best AI on the planet. DevOps team deployed the other 19.1% straight to production. One script opened port 22 to the internet. Three services down Monday morning. Eleven hours to roll back. #MilitaryAI #AIforGIs