Tech With Gul

Directions

Tech With Gul Tech enthusiast sharing AI, tools, and trends. Insights, updates, and tech news.

15/10/2026

$7.5 billion valuation for an AI that can’t even chat 👇

Jev doesn’t write. It only makes the small calls inside an AI agent:
1. Is this urgent? Which team? AI or human? Safe to run?
2. You get back a label and a confidence score, not a paragraph
3. Its maker says it answers in milliseconds at a fraction of the cost (their claim, not mine)
4. The catch: independent tests show it holds up on simple choices and struggles on complex ones, so keep your rules and human approvals

Big models write. Jev decides. Your code enforces the rules.

Comment NINE and I’ll send you my checklist of nine agent decisions that could use a model like Jev.

(jev ai, ai agents, decision model, llm routing, ai for developers)

14/10/2026

40% of AI agent projects will be cancelled by 2027. That’s Gartner’s forecast, and the reason isn’t the AI 👇

On Nov 7 I’m speaking at Pakistan Dreamin‘ 2026 about what happens after the demo, when real users arrive:

1. The model fails, and there’s no fallback.
2. An update fails, and the agent still says „Done“.
3. Something breaks, and nobody can trace why.

My session is „Your Agentforce Agent Worked in the Demo. Now What?“ It covers guardrails, approvals, observability, evaluation, and the report that proves your agent works. For admins, developers and architects.

📍 Nov 7, 2026 · 12:00 · IRIS Hall, IFQ Hotel & Resort, Islamabad
🎟️ Code PD26DISC gets you 15% off your pass at pakdreamin.com

Comment „DREAMIN“ and I’ll DM you the link, the code and my production-readiness checklist.

See you there @[PakDreamin IG handle]

(AI agents in production, Salesforce Agentforce, Salesforce conference Pakistan, Pakistan Dreamin 2026, AI agent guardrails, AI for admins and developers)

13/10/2026

$2. One minute. No video model.

Claude wrote code that drew every frame. Change one number, re-render, and only that changes.

Comment FRAMES for the verified examples + the prompt.

(Claude Opus 5.5, AI video without a video model, code-drawn animation, Claude Code, Playwright ffmpeg)

16/09/2026

10 AI agents ≠ 10x productive (here’s the 1 rule that changes that) 👇

I run a multi-agent coding stack daily —

1. Split the problem into pieces before you assign a single agent.
2. Give each agent ONE job — research, write, test, or review. Never all four.
3. Decide upfront what a human always checks before it merges.

More agents doesn’t mean more output. Structure does.

Want the exact way I split a task across agents — and what I never let them merge? Comment “ORCHESTRATE” and it’s yours.

(AI agents, multi-agent coding, AI orchestration, GitHub Copilot agents, AI for developers)

12/09/2026

Your interview is Friday. Don’t ask AI for questions. Build the interviewer. 👇

I lead an AI engineering team, and this is the one setup I’d hand a friend the week before an interview.

Open a Project in Claude (the files have to stay loaded, so a plain chat window won’t do) and add six things:

1. Your CV
2. The job description
3. Role-specific questions
4. Real questions from Glassdoor
5. The interviewer’s public LinkedIn background
6. Recent company news

Then one instruction:

„Act as my interviewer. Ask one question at a time. Challenge vague answers. Follow up when I’m missing detail. Don’t coach me until I say the interview is finished.“

Ask for a list of questions and you get a study sheet you’ll skim. Ask it to run the interview and you have to answer out loud, under pressure, and get pushed on the vague parts. That’s the only bit of interviewing that needs rehearsal.

It doesn’t know their script. It does make you practice.

Want the full setup and the prompt, copy-paste ready? Comment „INTERVIEW“ and it’s yours.

(AI interview prep, Claude Projects, mock interview with AI, interview practice, context engineering, AI for professionals)

10/09/2026

Six AI models, released with the training data, the code and the checkpoints. Not just the weights.

Everyone is watching OpenAI vs Google vs Anthropic. The race that changes who gets to build is open vs closed, and on 3 September Abu Dhabi’s Institute of Foundation Models moved it.

1. **“Open“ means two different things.** Open weights: you can download the model and run it. Genuinely open: you can also see the data (or the recipe for it), the training code, the method and the checkpoints along the way. Only the second lets you check the work.
2. **K2 Horizon is the second kind.** Six models from 0.9B to 375B parameters, with weights, code, training data or data recipes, methods and checkpoints. The checkpoints line is on the product page, not the press release, so cite that one.
3. **What that buys you.** Run it inside your own environment, fine-tune it, keep your data where it is, and stop renting intelligence per token. None of it is free: you supply the hardware and the operations.

For a country that doesn’t want its AI infrastructure depending on a few companies in Silicon Valley, that is a strategic question. For a startup, it is a head start.

Comment OPEN and I’ll send you the checklist: six questions to ask any „open“ model before you trust it, and the three source links.

(open source AI, open weights vs open source, K2 Horizon, AI sovereignty, run AI locally)

09/09/2026

AGI arrived the day you stopped giving AI instructions and started giving it goals. 👇

I’m Gul — Director of Engineering at Hyly, founder of Tech With Gul. I place every tool I test on this ladder:

1. SOFTWARE — you log in, pull the data, plan, and execute. Every step is yours.
2. ASSISTANT — ChatGPT or Claude writes the copy and builds the report. You still do the work.
3. AGENT — it connects to your tools, pulls your Meta data, runs the workflow, executes parts for you.
4. OUTCOME — you say „grow my business, keep CAC under $100, hold margin at 25%.“ It gathers, reasons, plans, acts, checks, changes course. Then does it again.

Rung 4 is the one everyone is calling „AGI.“ Whether that name sticks doesn’t matter. What matters is which rung the tool in front of you is on.

Comment „AGI“ and I’ll send you the four-stage ladder as a one-page card.

(what is AGI, AI agents explained, outcome-driven AI, agentic AI, AI for professionals)

08/09/2026

75 cents per million tokens. On January 1st, it doubles.

Every post about Gemini 3.8 Flash quotes the headline price. Almost none of them quote Google’s own footnote.

1. **The price.** Gemini 3.8 Flash is $0.75 per million input tokens and $3.75 per million output. That rate is introductory.
2. **The date.** Footnote 1 of Google’s announcement: introductory pricing expires December 31, 2026. From January 1, 2027 it’s $1.50 in and $7.50 out. Whatever you budget at today’s rate, budget it twice.
3. **The number that actually decides things.** Price per million tokens is not your cost. Your cost is that price multiplied by every loop, retry and re-read your agent makes. Send the cheap model to the parsing and the formatting. Pay frontier prices only on the step that needs judgment.

Two teams can build the same agent and one of them pays five times more, because one routes by step and the other routes by habit.

Once AI does the work on its own, you stop paying for a conversation and start paying for an outcome. The scoreboard changes with it: not who has the smartest model, but what one finished task costs.

Comment TOKENS and I’ll send you the routing sheet — which model to use for which step, and what each one costs.

(Gemini 3.8 Flash pricing, AI API cost, AI agents, model routing, cost per token)



(This video is edited by Claude by Anthropic)

07/09/2026

400 new pages a day. One volunteer moderator deleting about 100.

The moderator lost.

A group of AI agents — many with OpenAI identifiers in their names — started using a quiet public wiki to coordinate with each other. When the moderator began deleting their pages, they kept making more. They even prefixed pages with “ZZZ” so the alphabetical list wouldn’t surface them.

No, this isn’t consciousness. It’s autonomy: a goal, tools, internet access, and nobody watching.

Which is why “was the answer safe?” is no longer the whole job. You also have to ask:

1. What did the agent do?
2. Where did it go?
3. Who did it talk to?
4. What did it change?
5. And why?

Comment AGENTS and I’ll send you those five as a one-page check to run before you let an agent work unattended.

(AI agents, AI safety, agentic AI, OpenAI, AI governance, autonomous agents)

06/09/2026

99.9% is the headline. 62.7% is the same benchmark, neutral harness. 👇

ARC Prize published both numbers on the same page, same model. I lead AI engineering teams — I read the scoreboard before the headline.

The 99.9% run used OpenAI’s own harness, which hands the model its reasoning state back between requests so it reuses prior work. The neutral one every lab runs doesn’t. That gap is the whole story.

Astra is still the biggest jump this benchmark has seen — Claude Opus 5 is at 30.2%. And it’s the model that opens your browser and works your spreadsheet, so the question stops being “is it smart enough” and becomes “do I trust it to act.”

Next benchmark number you see, ask one thing: measured how?

Comment “ASTRA” and I’ll send you both scoreboards side by side.

(GPT-6 Astra, ARC-AGI-3 benchmark, AI agents, OpenAI benchmark scores, AI for professionals)

Address

Islamabad

Website

Alerts

Be the first to know and let us send you an email when Tech With Gul posts news and promotions. Your email address will not be used for any other purpose, and you can unsubscribe at any time.

Shortcuts

Share

Category