Ben Badejo (@BenjaminBadejo) header
Ben Badejo (@BenjaminBadejo) profile picture

AI Implementation for SMBs & Individuals đź“§[email protected]📱+1-212-203-7591 https://teapaw.org/t.co/A5I2CKJXRm • https://teapaw.org/t.co/GjhCxdvUYU • https://teapaw.org/t.co/DKmx0SSfL7 🇺🇸✡️🇮🇱

Tel Aviv / NYC
Joined October 2014
Photo by Ben Badejo (@BenjaminBadejo)Photo by Ben Badejo (@BenjaminBadejo)Photo by Ben Badejo (@BenjaminBadejo)Photo by Ben Badejo (@BenjaminBadejo)Photo by Ben Badejo (@BenjaminBadejo)Photo by Ben Badejo (@BenjaminBadejo)Photo by Ben Badejo (@BenjaminBadejo)Photo by Ben Badejo (@BenjaminBadejo)Photo by Ben Badejo (@BenjaminBadejo)Photo by Ben Badejo (@BenjaminBadejo)
Sightless will be available for direct download to your iPhone or iPad over Wi-Fi/cellular, without the App Store, for all iOS users, on Monday. It uses GPT-6/6.1 models and GPT Live / Realtime voice models to work in and across every single iOS app, and multiple apps to get complex tasks done, all by voice. The macOS companion app is only necessary if you want Codex / OpenClaw / Hermes / Claude agents to control your iOS device, rather than using your voice. This is everything that Siri should have been but will never be. And it’s way faster than it was in the video below from early October. And it’s done. It works. This is real.
Why use Muse or GrokBot when you can just use Sightless? You don’t have to connect it to everything because it uses your whole iPhone or iPad, which already has everything. 🤓
11
4
2
27
6,728
Ben Badejo retweeted
DeepSeek V4.1 Flash, fully local on one Mac Studio (M5 Ultra, 512 GB). 98 tok/s mixed prompts, 74 plain writing, 365 rewriting files. Still 69 tok/s at 1M context; the first read of 1M takes 9 min. @ddalcu's mlx-serve + 2 small patches
6
2
1
27
1,818
OpenAI is rolling out a free CSAM image classifier for approved organizations to detect CSAM material, whether in non-AI-generated images or in AI-generated images. If you don’t know what that acronym stands for, that’s a good thing. I know people are eager for new GPT models to be released, but if this is the sort of non-GPT OpenAI is also creating, it’s probably worth letting them finish their work on their timeline. I would not be surprised if this was developed in part because they realized it was necessary. Unfortunately, folks at frontier labs deal with real safety issues. This is one of them. The safety issue isn’t all theater.
2
6
715
GPT-6.1 Sol vs. DeepSeek 4.1 1017 Abliterated. This is a great example of why open-weight abliterated models are essential. For legal and product policy reasons, the leading closed-weight models just won't do certain things, like analyze and process copyrighted song lyrics. The first image is the result from GPT-6.1-Sol; the model would not provide, transliterate and translate the full song lyrics to fulfill my request. The following three images are the result from DeepSeek 4.1 1017 Abliterated; it handled everything flawlessly in one shot and without requiring me to carry out any steps manually.
5
3
2
24
2,866
Links: Local model download from HuggingFace, to run on your own hardware (note: this link is just one of many variants, by Obliteratus): huggingface.co/OBLITERATUS/D… Or if you don’t have the requisite hardware: API, via Refuseless: refuseless.com/models/deepse… API, via Venice: venice.ai/models/deepseek-v4…
175
Ben Badejo retweeted
GPT-6.1 Sol vs. DeepSeek 4.1 1017 Abliterated. This is a great example of why open-weight abliterated models are essential. For legal and product policy reasons, the leading closed-weight models just won't do certain things, like analyze and process copyrighted song lyrics. The first image is the result from GPT-6.1-Sol; the model would not provide, transliterate and translate the full song lyrics to fulfill my request. The following three images are the result from DeepSeek 4.1 1017 Abliterated; it handled everything flawlessly in one shot and without requiring me to carry out any steps manually.
5
3
2
24
2,866
Ben Badejo retweeted
we are living in the golden age of custom software
60
42
3
597
10,535
I told folks to get the M5U 256GB / 512GB and I was not wrong.
Improvements dropping daily.
1
6
964
OpenAI’s next model will basically blast us into outer space. It will be that good. It’s clear what happened: they skipped the model they wanted to release in favor of a better one. They are going to release the kraken.
3
1
2
21
1,984
And all of you doubters are going to look…
235
Ben Badejo retweeted
Sightless will be available for direct download to your iPhone or iPad over Wi-Fi/cellular, without the App Store, for all iOS users, on Monday. It uses GPT-6/6.1 models and GPT Live / Realtime voice models to work in and across every single iOS app, and multiple apps to get complex tasks done, all by voice. The macOS companion app is only necessary if you want Codex / OpenClaw / Hermes / Claude agents to control your iOS device, rather than using your voice. This is everything that Siri should have been but will never be. And it’s way faster than it was in the video below from early October. And it’s done. It works. This is real.
Why use Muse or GrokBot when you can just use Sightless? You don’t have to connect it to everything because it uses your whole iPhone or iPad, which already has everything. 🤓
11
4
2
27
6,728
Ben Badejo retweeted
First weekend playing with the new M5U 512GB. Spent it trying different models in some of my agentic workflows. I've been really impressed with DeepSeek-V4.1-Flash-1017. It's really close to a frontier feel. After using it for awhile I moved to the oQ4e-mtp version. It shrinks the N-gram table enough so it can fit in memory. Increased tok/s 40-50%, and I didn't notice any degradation in outputs. Uses about 405GB memory. Here is some benchmark data from oMLX (I'm also really enjoying oMLX, it's great in all the right places, I hope @jundotkim and @huggingface keep at it):
18
10
2
103
6,284
Tech people think too much about forests and not enough about trees Loving “humanity” while not caring about humans. Building for “future generations” while not raising children But the optimal amount to sacrifice rationality for sentiment is not zero Sometimes the best way to make the world better isn’t through future hypotheticals or abstract aggregates but solving specific problems, for real people, in the present
98
174
34
1,716
72,530
GPT-6.1-Sol has felt really rough today. Like, really rough. Not sure if it’s because of what I’m working on or something else. Even on ultrafast, things are taking a very long time…
9
15
2,173
Ben Badejo retweeted
DeepSeek's official 475 GB DSV4.1 Flash is doing 90 decode (up from 70) using TensorFold on M5 Ultra! We're starting to squeeze Mac's new silicons, lets see where it takes us.
These M5U's move 🤯🚀 DeepSeek’s official DSV4.1-Flash-1017 weights are about 475 GB, served at temperature 1.0 and top_p 1.0, with context capped at 131k for testing. Single stream 1,300 prefill 70 decode This is usable frontier intelligence on your desk. 🫡
9
3
2
95
9,764
Ben Badejo retweeted
🚨 Original work alert 🚨 (from my wife) I actually don't know a good solution for this problem. It may or may not be solvable Creating benchmarks and RL tasks for TRADING might prove to be quite a bit of a challenge, at least when it comes to measuring performance on public historical data Imagine that you feed 2025 prices to an LLM day by day without any other context, for backtesting It turns out most LLMs have memorized copper prices out of all things and they can just cheat... It is likely that they have memorized the prices and market events of most significant things, so you cannot just create a task based on public information What would be a good way to solve this? Could such market knowledge be ablated during training, and added back at the end? Would that be practical at all? Or can we create synthetic market data reliably? Or do we only have 3-6 months back to the pretraining cutoff date? What about training in real time 🤔 Some of these questions likely already have answers since deepseek literally is an offshoot from a fund and has a long history. But I haven't seen anyone do communication on this like the way @qrlow does. So do give her a follow!
It's really hard to create a trading benchmark. As I found out, AI models are really good at cheating. On the surface, their PnL looks great. But that's only because they time travel and use information that could not have been known. This was what happened when I made AI play a simulation game initially modelled after 2025 events and prices. Some models appeared to occasionally have more morals than others. Opus for example, could but initially decided not to cheat. Though in subsequent games it also decided to cheat anyways, lol. In any case, when the models couldn't cheat, PnL plummeted To me, this simply shows that they were not trained for the trading task. So it's actually a big opportunity for whoever can train an "AlphaTrade" model first (just like how "AlphaGo" was trained for the game of Go) But how do we even start training a model on this? Model training starts with data, and the best data for this sits at trading houses. But there is no way trading houses will share any of that. This is where synthetic data comes in. If there can be a simulated world where events and corresponding market variables can be realistically modelled, it becomes an ideal sandbox for reinforcement learning. I created a Synthetic Year Generator engine as a means to achieve such a simulated world. We start off with a set list (a menu) of possible events, each with possible paths and corresponding impacts on prices and other market variable for each path. Events can then be pieced together to rapidly form new synthetic years, with events compounding or reducing each others effects as the events feed into a central underlying supply and demand model. Actually, when I first started creating synthetic years, I was spending a lot of time "handcrafting" them. I still think if a synthetic year was designed by a human expert in an "artisanal" way, it would be higher quality than what is generated from the engine. The engine does speed up the synthetic year creation process, but does not omit the need for a human judgement from a subject matter expert. Full project on my blog: qrlow.me/blog/tradingdesk-si… Short intro video:
1
1
7
906
If you use Omarchy, please tell me succinctly what the benefits of it are for you personally, and why you like it so much.
13
15
3,310
Note: I haven’t tried it yet — trying to get feedback before I do.
1
455
Is the latest version of the macOS ChatGPT desktop app constantly restarting for anyone (and then hanging on the loading screen)? On version 26.1111.21308. I made sure Codex CLI / daemon are both up to date but the issue persists.
1,451
Ben Badejo retweeted
If G-d enables me to make real money from something that will help so many people…wow. Every person with a physical disability but who can use their voice will be able to have their entire iOS device operated by AI, live, in real time, in every app, 24/7. Please, G-d. TYIA.
Sightless will be available for direct download to your iPhone or iPad over Wi-Fi/cellular, without the App Store, for all iOS users, on Monday. It uses GPT-6/6.1 models and GPT Live / Realtime voice models to work in and across every single iOS app, and multiple apps to get complex tasks done, all by voice. The macOS companion app is only necessary if you want Codex / OpenClaw / Hermes / Claude agents to control your iOS device, rather than using your voice. This is everything that Siri should have been but will never be. And it’s way faster than it was in the video below from early October. And it’s done. It works. This is real.
7
3
40
1,875