Yosef Shammout

Building with graphs, agents & LLMs. SWE @ Ford. CS @ Wayne State. UFC fan

[Yosef Shammout @yosefcodes]

“my doctor” and “my ex” are both people you know. You’d react very differently to a text from each. Most AI wouldn’t.

It flattens every relationship into the same type. Connection is connection. That’s where it loses the plot.

• The fix is letting the model learn different rules for different kinds of connections, the way you do without thinking about it.
• This matters anywhere relationships carry meaning: medical records, fraud detection, recommendations, social networks. Basically everywhere.
• The strangest part: it often helps most on the rare cases, the ones normal models give up on.

More soon.

[Yosef Shammout @yosefcodes]

@Baxate yea microsoft did this with MAI-Thinking-1, zero distillation, and it beat sonnet 4.6 in their blind preference tests

[Yosef Shammout @yosefcodes]

Two weeks out, Justin Gaethje is bringing it home on the White House lawn.

[Yosef Shammout @yosefcodes]

how do you decide when a model should think longer on a given input? confidence, difficulty, something learned? curious what’s actually working for people.

#AI #machinelearning

[Yosef Shammout @yosefcodes]

still kind of wild that anthropic is paying spacex $1.25 billion a month to train on colossus. you don’t rent a competitor’s data center unless it’s the best one on the planet. people really sleep on how far ahead xai’s compute is.

[Yosef Shammout @yosefcodes]

@PalmerLuckey @Lazarus_Capital @markpoloncarz UFC is peak, its takes incredible talent, skill, and discipline (maybe not health discipline for prates 😂). UFC this summer is STACKED

[Yosef Shammout @yosefcodes]

@Baxate Mexican coke > “regular coke”

[Yosef Shammout @yosefcodes]

love this, the “more than did it finish” part is where i landed too. built mirage for it: a deterministic policy gate for agent actions, rules in yaml, no llm in the decision loop so it can actually fail a ci build. how are you scoping steeringbench, workflow discipline or the agent’s actions mid-run? link’s in my bio if useful

[Yosef Shammout @yosefcodes]

opus 4.8 beats gpt-5.5 and gemini 3.1 at the same price as 4.7. better model, flat price. that’s a pattern worth watching

[Yosef Shammout @yosefcodes]

@NaderLikeLadder the prefix is just a timestamp. eHarmony, iPhone, OpenAI, and now the agent wave. AgentForce, AgentKit, watch the naming converge

[Yosef Shammout @yosefcodes]

@elonmusk Huge for Grok 5. Does this stack make it easier to experiment with even larger models or novel architectures? The >10x throughput potential vs JAX changes everything at Colossus scale

[Yosef Shammout @yosefcodes]

true fault tolerance in action. starship lost one of six engines mid-flight and still completed its mission, deploying payloads and surviving reentry. peak engineering. congrats @SpaceX and @elonmusk on flight 12. #Starship

[Yosef Shammout @yosefcodes]

Been thinking about this lately.

The best agents aren’t always the smartest ones. It’s more about solid memory that doesn’t get messy and good brakes so they don’t spiral.

Feels like we’re finally moving past hype into stuff you can actually trust.

Anyone else seeing the same thing? What’s one small change that helped your agents?

[Yosef Shammout @yosefcodes]

@BipolarBastard the interesting case is structured data, things like medical records, transaction graphs, social networks. models there can’t tell one kind of connection from another unless you build it in. llms pick this up from language, which is a different problem entirely

[Yosef Shammout @yosefcodes]

“my doctor” and “my ex” are both people you know. You’d react very differently to a text from each. Most AI wouldn’t.

It flattens every relationship into the same type. Connection is connection. That’s where it loses the plot.

• The fix is letting the model learn different rules for different kinds of connections, the way you do without thinking about it.
• This matters anywhere relationships carry meaning: medical records, fraud detection, recommendations, social networks. Basically everywhere.
• The strangest part: it often helps most on the rare cases, the ones normal models give up on.

More soon.

[Andrej Karpathy @karpathy]

Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.

[Yosef Shammout @yosefcodes]

self-hosted sandboxes and mcp tunnels lock down the runtime well. but the model still picks which tools to call and what params to send. if the tool defs are too open, it can still try a broad delete or update inside your vpc. the sandbox contains the mess but doesnt stop the bad idea upfront. Im wonder how strict the mcp is on the checking calls before they run?

[Yosef Shammout @yosefcodes]

I've noticed that seasoned SWEs don’t one-shot with their agents.

They have a very specific set of instructions for a specific job.

[Yosef Shammout @yosefcodes]

Using @grok is genuinely nice.

Straight to the point, zero filler, no lecture.
Just the signal I need when I’m knee-deep in research.

The rest of the ecosystem could learn from this.

[Yosef Shammout @yosefcodes]

capability is not the hard part of agents anymore. control is.

[Yosef Shammout @yosefcodes]

agent eval benchmarks average over every tool call.

that buries the one call you actually care about: the rare, high-consequence one.

averages lie. i learned that benchmarking medication prediction.