Phil Youssef

Software Engineer | Builder 🏗 & 🍝 pasta enthusiast | 🇺🇸 🇨🇦 🇪🇬 🇮🇹

[Phil Youssef @PhilYoussef]

These minimum viable delights, have minimum viable ends.

[Phil Youssef @PhilYoussef]

Opus 5.5 vs GPT 6 Sol
Opus is ~20% smarter but costs ~5x more & produces 4x more tokens

Use Opus for non time sensitive planning & Sol as a daily driver?

https://t.co/UKJKInUSAA

[Phil Youssef @PhilYoussef]

@iamlemec It's quite real; I've been playing around with it for the last day. I think they've invented a new type of AI that will compliment LLMs in many use cases.

[Phil Youssef @PhilYoussef]

@mycreditjourney https://t.co/Zep6DnuRds -> Join Waitlist
Got access same day...

[Phil Youssef @PhilYoussef]

Got access to TypeSafe. Asked Astra to wire Jev up to StarCraft.
It one-shot the harness and won the mission.

[Phil Youssef @PhilYoussef]

@kimmonismus seems to be able to do starcraft passably https://t.co/qkvRiR59hN

[Phil Youssef @PhilYoussef]

Got access to TypeSafe. Asked Astra to wire Jev up to StarCraft.
It one-shot the harness and won the mission.

[Phil Youssef @PhilYoussef]

@dotpem @motatoeshq can confirm - i was able to replicate with starcraft: https://t.co/qkvRiR59hN

good job @typesafeai !

[Phil Youssef @PhilYoussef]

Got access to TypeSafe. Asked Astra to wire Jev up to StarCraft.
It one-shot the harness and won the mission.

[Phil Youssef @PhilYoussef]

@ziwenxu_ It can also play Starcraft somewhat competently too. $0.39 to beat a mission: https://t.co/qkvRiR59hN

[Phil Youssef @PhilYoussef]

Got access to TypeSafe. Asked Astra to wire Jev up to StarCraft.
It one-shot the harness and won the mission.

[Phil Youssef @PhilYoussef]

The winning run cost about $0.40 in Jev API usage.

- Input: 9,445,640 tokens × $0.042/million = $0.3967
- Output: 234,062 tokens — free

[Phil Youssef @PhilYoussef]

Details for the curious:
Original 1998 StarCraft shareware, running in the browser via BottleShip. Jev gets structured game state (own units, visible enemies, economy), picks a command, and the harness executes it with real mouse clicks and hotkeys. Game pauses during inference.
Winning run: 421 decisions, 17.5 min, ~9.4M input tokens, 383ms median latency.
Took 16 attempts to get there — most failures were harness bugs, not the model. Every one is documented.
Repo: https://t.co/8lMcVwAScp
8× video + full 1× + evidence bundle: https://t.co/HkqECmFogD

[Phil Youssef @PhilYoussef]

@jacobandreou what happens in the fold stays in the fold

[Phil Youssef @PhilYoussef]

@jeff_weinstein @stripe @link wouldn’t happen to have something i can try out?
I’ve got some time between jobs coming up and need to make https://t.co/kMPOtAs8y7 agentic… 😏

[Phil Youssef @PhilYoussef]

@natolambert you overestimate these companies capacity to build controls in lock step with their scale. most companies i’ve seen moving this fast are like having an adult head in a baby’s body.

[Phil Youssef @PhilYoussef]

@finbarr is your orchestrator on a loop of some kind?

[Phil Youssef @PhilYoussef]

@sriramk really curious if blender is working so well because OAI had an RL gym for it set up as part of post training

[Phil Youssef @PhilYoussef]

build a 2 stage data pipeline using python scripts. stage 1 extracts metadata from individual trajectories based on types of things you’re interested in. index stage 1 data. then stage 2 looks for patterns across trajectories to generate a report. you need to have an idea what you’re looking for - but this approach scales well to a few orders of magnitude bigger than your data set.

[Phil Youssef @PhilYoussef]

@Suhail i hit this regularly too - going to give your skill a spin

[Phil Youssef @PhilYoussef]

@stevehou ask it to create an html based slide deck. powerpoint as a format is quite limited compared to full html

[Phil Youssef @PhilYoussef]

😮‍💨

[Phil Youssef @PhilYoussef]

6 Astra Ultra Fast

[Phil Youssef @PhilYoussef]

@meganeoleary big bufo fan over here as well … check out https://t.co/kMPOtAs8y7