
Venture Builder, Data Scientist, Financial Engineer | Ex Director of Engineering & AI @AmericanExpress |@Cornell MBA/MSc
Connect with me ➡️
Joined July 2010
- Tweets2.6K
- Following5.5K
- Followers34.2K
- Likes24.8K
Pinned Tweet
Introducing the next version of Son of Basel, powered by @FirmAdaptAI
This is a completely custom-designed and engineered clone of me, managing a team of digital employees who work 24/7/365, always available and infinitely scalable as needed.
I have Son of Basel join me for all my meetings and call him directly to brainstorm and get progress updates.
This took FOREVER for us to build, and I am really proud to be sharing it
There is absolutely nothing like it on the market as it's my clone!
Using a proprietary, finely tuned, and calibrated set of models designed to mirror me
Core Automation trained one AI model in two data centers 9,000 miles apart, and it came out about 1% better than training in one place on the same compute.
The US and Asian sites only had to exchange updates every 50 training steps. I like these types of tests.
3
8
967
AI coding agents can write code that runs and still fail at the real job. In a Sejong University test, Claude Code with Sonnet 4.6 wrote robot training code that ran 9 times in 10...
...yet none of it got a simulated robot to walk.
Real-life production tests are critical!

ALT Diagram of the ARPG framework (Figure 2): in a modeling phase, agents for the objective, variables, constraints and rewards build the robot's task design while an Error Correction Agent checks it, and only then do seven implementation agents and a Coding Agent write and debug the Isaac Sim code.
13
6
901
Read more here arxiv.org/abs/2611.04974
170
ChatGPT's top product pick changes with the order it sees the links in. The same rewritten pages came first 74% of the time in one link order and 94% in the reverse order.
1
3
835
University of Illinois researchers tested it with 80 made-up brands as part of a GEO experiment. I like these in the wild tests because we don't know exactly why some content ranks higher than other content
1
193
In an Imperial College London test, simple fixed rules beat a second AI at catching a small AI's bad moves in a simulated chemical plant.
With rule checks, Meta's Llama 3.2 3B shut the plant down in 7/20 faults, against 15 when another AI checked its work, pretty interesting
671
Claude Opus 5 aced the individual tasks of one project, averaging 88 out of 100, but the finished work it handed in scored only 55. It still had the top score on DeliverBench, a new test built from real professional work like financial due diligence.
1
1
753
Microsoft built Copilot Cowork and launched it worldwide in four and a half months, with AI agents writing most of the code. On one day in October, nine engineers merged 1,211 code changes into it.
Pretty impressive!
7
5
784
A team of AI agents trained to split up the work finished 4 times faster than a single agent on the same budget. General Reasoning trained teams of up to 9 agents that share one reward, so they learn to work in parallel.
3
5
766
Basel Ismail retweeted
One year in crypto (Oct 19 to Oct 19):
Bitcoin -25%
Ether -40%
Solana -44%
S&P 500 +15%
At their lowest point in that year, Bitcoin was down 49% from its high, Ether 62% and Solana 69%.
Basel Ismail retweeted
One year, three assets:
Apple +29.5%
Nvidia +18.5%
Bitcoin -24.6%
S&P 500 +14.5%
Over the same year Nvidia insiders sold about $2.0B of stock in 240 sales, and made no open-market buys.
Basel Ismail retweeted
GameStop CEO Ryan Cohen has bought GME on the open market five times since Sep 10.
About $97M in total. The latest: $22.8M on Oct 13 at $26.98 a share.
He now holds 42.49M shares, worth about $1.12B.
Basel Ismail retweeted
Quick Introduction to Prediction Alpha
World's Only Prediction Market Mispricer, Whale Watcher, and Arbitrageur
Basel Ismail retweeted
Quick update on the open trades posted to Trade Alpha across multiple asset classes by @BaselIsmail
Source: blockcircle.com/trade-alpha











