AI developer | Cybersecurity researcher | Technical writer | Exploring the systems shaping AI, security, and digital policy.
Nigeria
Joined January 2018
- Tweets239
- Following600
- Followers104
- Likes167
Muhammad Abdullahi dattijo retweeted
Best 15 accounts to follow in AI:
@karpathy = LLMs king
@steipete = built openclaw
@gregisenberg = startup ideas king
@rileybrown= vibecode king
@jackfriks = solo apps king
@levelsio= startups king
@marclou = startups king
@EXM7777 = AI ops + systems king
@eptwts = AI money twitter king
@ai_explorer25= AI queen
@godofprompt =prompt king
@vasuman = AI agents king
@AmirMushich = AI ads king
@0xROAS = AI UGCs king
@egeberkina = AI images king
Follow them all and learn.
Muhammad Abdullahi dattijo retweeted
How well do AI image judges perform, and what are the areas where they still struggle?
@DattijoMuhammad explores a complex set of challenges in his new deep dive on vision AI models. towardsdatascience.com/i-bui…
Caveats: small test, artificial differences, Gemini 3.1 Flash Lite instead of the main Gemini model (quota), and not comparable to VisionQ scores. Full article: towardsdatascience.com/i-bui…
GPT-4o fell for a different one. On the tilt question, where neither image was tilted more and the right answer was SAME, it picked an image 72.5% of the time, often sounding confident. Claude did that 17.5% of the time, Gemini 28.3%.
The different trap: Claude answered B in 23 of 24 replies on the compression question and was right only 11 of 24 times. Across the swapped image orders, it gave the same letter both times in 15% of cases. A position habit, not a judgment.
On plain pairs they were strong. Asked which image is better: GPT-4o 97.9% correct, Claude 92.7%, Gemini 100%. The vivid but blurry trap fooled none of them: 0% for all three.
Setup: 12 original AI generated images, 5 pairs each (blur, grain, compression, darkening, plus the trap), so 60 pairs. Three questions per pair, each asked in both image orders: 1,080 questions. Judges: GPT-4o, Claude Sonnet 5 and Gemini 3.1 Flash Lite.
Retrieval completed 88.9% of tasks on mini and 85.6% on Haiku, and fetched every needed piece in 93% of runs. Caveat: 30 tasks, one made-up person, so read it as what to test, not a verdict. Full article: towardsdatascience.com/i-tes…
Full history also struggled with corrected facts: mini completed 46% of tasks on corrected facts vs 66% on unchanged ones. The running summary lost facts too: mini kept the newest value for 20 of 30, Haiku for 28.
The four designs: no memory, full history, a running summary (about 1,000 words), and retrieval (184 chunks, the 5 closest go into each prompt). No memory scored 0% on tasks, as expected.
Setup: a made-up project manager with 80 history entries over a year (about 40,000 tokens), 11 of them correcting an earlier fact. 30 tasks that end in a tool call, 30 plain questions. GPT-4o mini and Claude Haiku 4.5, 1,260 runs logged to W&B.
I tested four ways to give an AI assistant a memory. The obvious one, pasting the whole history into every prompt, was the worst value: 20 to 31 times the cost of retrieval, and it did not win on accuracy either. Thread:
Final setup: a cheaper model with the strong model as backup. 129 calls, $0.008, all 101 matches, zero false positives. Caveats: one run per setup, and I picked the escalation threshold after seeing results. Full write-up: towardsdatascience.com/can-a…
Plot twist: giving the strong model LESS text helped. On one hard listing, the full listing prompt got it right 0 of 10 times. Title and body only got it right 10 of 10.
Step 2: check the pets field in code instead of asking the model. 157 calls, $0.11. This was the biggest single drop, 44% of the total savings.
Step 1: a plain database query on city, bedrooms and rent. No model needed. Calls dropped from 2,500 to 264 and cost to $0.20, with the same matches.
Setup: 500 rental listings from the UCI dataset, 5 requirements each, so 2,500 listing and requirement pairs. Sending all of them to the strong model is 2,500 calls and about $1.88. I traced every call with W&B Weave and computed cost from token counts.