Ali Shohadaee

AI Engineer | I use bitter coffee as a weapon to write about AI and what I learn building it

[Ali Shohadaee @alishohadaee]

enterprise AI agent buyers don't want more autonomy. they want to see what the agent did

G2 went through 3,222 reviews of AI agents from companies with 1,000+ employees (231 products)

trust is already high:
> 77% trust their agent quite a bit or very much
> 53% use it most of the time or every time

but autonomy still isn't the ask:
> 60% want guardrails, 21% limited autonomy
> 7% want total autonomy
> even among those who trust it very much, only 15% would let it run alone

asked what would raise trust, 79% picked visibility or control:
> explain its decisions: 32%
> more guardrail settings: 29%
> warn before questionable actions: 18%
> better reliability: 21%

and watching isn't free. in G2's separate agent survey, 45% say oversight and rework eat more than a quarter of the gains

[Ali Shohadaee @alishohadaee]

@sertoken888 ☕️ + AI💪🏻

[Ali Shohadaee @alishohadaee]

🫵 hey
I share personal thoughts about AI and what's happening

would love to connect with AI builders and enthusiasts 🍻

[Ali Shohadaee @alishohadaee]

@BennidaB sure, happy to🍻

[Ali Shohadaee @alishohadaee]

@alex__nedelea Glad to connect 🍻

[Ali Shohadaee @alishohadaee]

@markjgg Happy to connect 🍻

[Ali Shohadaee @alishohadaee]

🫵 hey
I share personal thoughts about AI and what's happening

would love to connect with AI builders and enthusiasts 🍻

[Ali Shohadaee @alishohadaee]

@Themassobrand Hey Masso, yeah man🍻

[Ali Shohadaee @alishohadaee]

@TarunShipIt hey Tarun, glad to connect 🍻

[Ali Shohadaee @alishohadaee]

@Jay36_9 hey Sanjay, happy to connect 🍻

[Ali Shohadaee @alishohadaee]

the good thing about all the AI thing is that

< as a student,
you’re sure about the ROL (return of learning)
< as a startup,
you’re sure about the ROI (hopefully 😂)

As a big corporate you have no choice but spend more and more hoping you’ll reach AGI first :)

[Ali Shohadaee @alishohadaee]

@niuniu_dev Hey Sean, pleased to know you🍻

[Ali Shohadaee @alishohadaee]

@dantegaleazzi Hey Dante,
I’ve been working on synthetic data generation for specific cases where almost-no data is available

[Ali Shohadaee @alishohadaee]

@vincenzo_naked Hey Vincenzo, nice to meet you 🍻

[Ali Shohadaee @alishohadaee]

@akacoding Connected buddy 🍻

[Ali Shohadaee @alishohadaee]

@quietforgelab connected 🍻

[Ali Shohadaee @alishohadaee]

and this complements it perfectly

[Matt Dancho (Business Science) @mdancho84]

This 277-page PDF unlocks the secrets of Large Language Models.

Here's what's inside: 🧵

[Ali Shohadaee @alishohadaee]

the most expensive tokens in your agent run are the ones it writes deep into a long context. YOU DONT NEED LONG CONTEXT.

"You don't need long context" (Belcak, Molchanov et al., Nov 3) puts numbers on it:
> serving at 256K instead of 16K costs 3.6x to 7x more per generated token (GLM 5.2, Kimi K2.6, Qwen 3 Coder Next, from H100 throughput)
> their fix is compaction: when context crosses a threshold, the agent writes down the task, its plan and the artifacts it still needs, then restarts fresh from that
> long documents stay as files the agent greps and reads in pieces. on long-input benchmarks that scored up to 40 points higher than pasting the whole input
> the extreme version, one action per context: Opus 4.5 hit 70.2% on SWE-bench Verified vs 70.0% with full history, with peak context down from 130–200K to 15–37K tokens

estimated savings: 2.2x to 8.8x on long agent runs for Opus 4.5

the honest part: costs are estimated from throughput and list prices, not real bills. compaction is lossy, so a task that needs an exact early detail fails unless that history stays readable. and open models break: Qwen 3 Coder Next fell from 53.0% to 3.6% on SWE-bench Verified at the tightest setting

so the lesson isn't "shrink your window". context is a cost line you design, not a limit you fill

if you run long agent sessions, log context size per step and try a compaction threshold before paying for the bigger window

[Ali Shohadaee @alishohadaee]

@latentspacepod @aiDotEngineer @Ninggao88 @swyx the headshot example from the Kill My SaaS writeup is the one I'd show anyone grading agent-built apps. the upload saved fine, but the public speaker page read from a different place, so only an end-to-end run as the speaker would catch it

[Shweta Kukreja @ShwetaKukreja_]

Canva is laying off 22% of its workforce which is 1271 employees, and this is what their CEO said.

We need more founders like thissss.

[Ali Shohadaee @alishohadaee]

@ShwetaKukreja_ I thinks that’s all the good a company can do for their employees in this moment, good job @canva

@Microsoft interesting, no?