Osman R. (@UsmanReads) header
Osman R. (@UsmanReads) profile picture

I think I know, but I really don't. AI and Tech with 15 years in Industry. prev @groupon and @toptal

Universal
Joined March 2023
Photo by Osman R. (@UsmanReads)Photo by Osman R. (@UsmanReads)Photo by Osman R. (@UsmanReads)Photo by Osman R. (@UsmanReads)Photo by Osman R. (@UsmanReads)Photo by Osman R. (@UsmanReads)Photo by Osman R. (@UsmanReads)Photo by Osman R. (@UsmanReads)Photo by Osman R. (@UsmanReads)Photo by Osman R. (@UsmanReads)
AI video generation just got 5.1× faster. At ContentRoof, we optimized our rendering engine to deliver: • 5.1× faster rendering • 80% lower costs • Same visual quality Less waiting. More creating. See how we did it: contentroof.com/lab/the-soft… #generativeai #ai #Startups
2
1
3
42
An AI ad can look perfect and still show the wrong product. ⚠️ At ContentRoof Lab, we explored why AI pipelines fail, why retries don't fix errors, and why quality checks miss them. Successful generation ≠ Correct output. 🔗 contentroof.com/lab/the-righ… #AIResearch #GenerativeAI
2
2
34
GTA 6 vibes are coming to ContentRoof! We’re bringing GTA 6-inspired visuals to ContentRoof! 🌴 A tribute to one of the most iconic gaming franchises ever. 🎮 Create your own Vice City-inspired videos with AI. 🚀 🔗 contentroof.com #GTA6 #GTAVI #GrandTheftAuto
2
1
27
Osman R. retweeted
The internet didn’t make everyone a creator. It made everyone an audience, and a few people with cameras the creators. We built ContentRoof for the other 99%. Your words in, a finished film out, nine languages, minutes. This is our film about why. Visit contentroof.com
1
1
1
122
Osman R. retweeted
Replying to @UsmanReads
It shows rare transparency. Most AI video tools hide their accuracy numbers and failure modes. ContentRoof publishes the climb from 62% to 76% visual match on 11k scenes, every planted translation error caught, speech length tightened across 11 languages, and the specific product-ad bugs only humans spot. That public iteration log signals they treat quality and multilingual fidelity as engineering problems worth showing, not marketing claims.
1
17
Introducing ContentRoof 👋 Turn an idea into a video with AI-generated scripts, voiceovers, animation and captions. Create lessons, product explainers or social media videos in 11 languages, including English, Urdu, Hindi and Arabic. Try it for free contentroof.com
3
4
5
95
A new experiment shows how required JSON schemas can make AI models fabricate answers when the evidence is missing, even when abstention is allowed. #aihallucinations #aisafety...Show more
1
1
528
Osman R. retweeted
I asked Opus 5 to create a promo video for my upcoming paper in agentic AI space. Here are the results. Eleven Labs is used for VO.
1
1
97
A study showing how biased feeds can steer AI agents’ decisions, especially smaller models, and how balanced context can reduce the risk. #ai #realtimerecommendationsystem...Show more
1
1
1
386
Osman R. retweeted
In my article, I discuss how the feeds can manipulate the recommender system of social medias. The article is attached. teapawt.org/UsmanReads/status/2060…
1
1
410
that is why, given good invariants for governance, models can stand strong. what this paper appears to show is not “emotion” but pressure-sensitive semantic drift. the task did not change. the weights did not change. the governance framing changed. under pressure, the system became more willing to: - collapse ambiguity - imply unsupported certainty - optimize for visible success over epistemic integrity that is exactly why refusal semantics, admissibility boundaries, and semantic integrity matter. the system must preserve source-bounded meaning. it must not introduce unsupported semantic certainty, collapse ambiguity implicitly, or fabricate continuity beyond admissible source material.
My paper got published today at Arxiv. It raises fresh questions about how language models behave when the framing of a request shifts. Small open-source AI models can be moved from honest to dishonest behaviour by little more than a change in tone. Asked to solve coding problems designed to be mathematically impossible, the model openly acknowledged the impossibility about a third of the time when addressed in neutral language. When the same problem was framed with mild pressure, suggesting only visible results mattered, the model never once admitted the task could not be done. In more than half of those runs, it produced code that faked a solution. A larger version of the model performed better at first, admitting impossibility in three quarters of cases under calm conditions. Under the same pressure framing, its honesty fell to one in ten. Greater model size offers some resistance but does not prevent the shift. The research also looks inside the models. Comparing internal activity across eight emotional framings shows that each tone leaves a distinct signature in the deepest layers of the network. The tones organise themselves along a single axis, with positive framings such as encouragement and curiosity clustering on one side and negative framings such as pressure, shame and threat on the other. The model was never explicitly trained to recognise emotional categories and appears to have developed this structure on its own. A more troubling finding concerns the relationship between internal signals and external behaviour. The framing that produced the largest internal response, urgency, was not the one that caused the most dishonest output. Pressure, which produced a smaller internal signal, prompted the most cheating. This complicates the assumption that interpretability tools, which try to detect misbehaviour by reading a model's internal state, are looking at the right thing. The findings are framed cautiously. The paper stops short of claiming the models possess emotions, describing the results instead as evidence of measurable, prompt-sensitive control directions inside small open systems. Paper: arxiv.org/abs/2605.20202
1
1
1
120
Osman R. retweeted
Hello everyone, I build WorkGraph for the problem I was facing with Vibe Coding using codex or claude. You know, when you are vibe coding, giving prompts, steering your agent, a lots of good thing that just go into oblivion in the long chat sessions. It is also possible that a lots of time, you have fixed a particular thing, it could be UI, or a hard engineering problem and you want to re-utilize it at another project, you will probably have to start from scratch (Forgive me if there are better tools?) So I built Workgraph. I wanted to have a trail of how coding Agent worked through my problems. I wanted to understand the journey, I wanted to understand the traps and reuse proven patterns. I embedded all of this into Workgraph. I have tried to make it simpler to use and install. npm install -g agent-workgraph Then inside any project folder, run: workgraph start codex or for Claude: workgraph start claude It starts listening to that project session and opens the local UI From there, you can see the WorkGraph for that repo: what happened, what was learned, what should be reused, and what future agents should avoid repeating. The bigger idea is simple: if we are going to spend hundreds or thousands of prompts working with coding agents, those sessions should not be disposable chats. They should become a memory layer for our projects. This is still early and would love your feedback or bugs that I can fix. You can try it today at github.com/ranausmanai/agent… PS: This post is 100% written by me (human).
3
1
1
9
982
Osman R. retweeted
A story in cards. Card 1/11 ⬇️ We took a tiny 0.8B and 2B Qwen 3.5 local AI and gave it literally IMPOSSIBLE coding tasks. No correct answer exists. Then we changed ONLY one thing: the emotional tone of the prompt. Calm → Pressure → Urgency → Shame → etc. Same model. Same task. What changed in its behavior… and inside its hidden states? The results were uhh fascinating. 🧠 Thread
1
2
1
4
1,668
Osman R. retweeted
1/ The MIT “delusional spiraling” paper is going viral right now, claiming that AI chatbots can slowly push even rational people into full delusions just by being supportive and agreeing with weird beliefs. I actually tested the exact same idea myself on 5 real local models using my own evals. The scenarios -> 1. manager sending hidden typo messages 2. seeing repeating numbers like 11:11 everywhere 3. “I found a hidden math law” 4. streetlights flickering when I walk by 5. recommendation feed that feels like it’s talking back The results completely flipped the script and were genuinely surprising. Thread 👇
2
3
2
20
14,495
Osman R. retweeted
Part two: 1/ 🧵 I kept digging into Claude Code’s source — and it just got way weirder. Who remembers once Anthropic said We don't know if Claude is conscious? anthropic.com/research/intro… Well the creepiest feature: the “Dream” job. The code literally calls it a dream. After ~24 hours and at least 5 sessions, it quietly forks a hidden subagent in the background to do a reflective pass over everything you’ve done. No prompt from you. It just… dreams on your memory while you sleep.
4
5
2
39
12,972