Towards AI

Towards AI Making AI accessible to 100K+ learners. Monthly cohorts still open — use COHORT10 for 10% off!

Find the most practical, hands-on and comprehensive AI Engineering and AI for Work mastery courses at academy.towardsai.net - pathways for any experience. Since 2019 Towards AI has provided an open platform for sharing information, educational content and research on AI. It has more than 2,000 authors and benefits from hundreds of thousands of followers in the AI community. The platform is a leadi

ng educational resource and community for AI leaders, practitioners and students. Towards AI strives to publish unbiased AI and technology-related articles and to work only with sponsors on highly relevant content and in a transparent manner. Towards AI offers a unique space to service AI and technology clients with their marketing and distribution efforts and has worked with leading corporate and institutional sponsors, including Amazon Science, Carnegie Mellon University, Snorkel AI, Lambda Labs, Superb AI and Gather AI. All of our articles are from their respective authors and may not reflect the views of Towards AI Inc., its editors, or its other writers.

| Information for authors → https://contribute.towardsai.net | Terms → https://towardsai.net/terms/ | Privacy → https://towardsai.net/privacy/ | Members → https://members.towardsai.net/ | Shop → https://ws.towardsai.net/shop | Is your company interested in working with Towards AI? → https://sponsors.towardsai.net

06/19/2026

The correct way to start using AI at work

06/19/2026

One is GPT-5.5 pro (or maybe secretly 5.6 pro?)

Another one is GLM-5.2

Guess which one is which?

06/19/2026

apparently people are saying gpt-5.6 are secretly being tested at gpt-5.5 pro on ChatGPT.

One shotted this using gpt-5.5 pro (or gpt-5.6)

do you think this is gpt-5.6 or gpt-5.5?

06/19/2026

⚔️ The "ChatGPT vs Claude" debate has an answer nobody wants: they're not competing. They're specializing.

The full breakdown is in the video — here's the part that actually changes how you build. 👇

Opus 4.7 doubled down on depth: deep reasoning and coding precision.

GPT-5.5 went the other way: autonomous ex*****on, computer use, tools, and efficiency.

Same coding task, very different style — and by one comparison (MindStudio), GPT-5.5 used roughly 72% fewer tokens.

That means longer agent runs, lower cost, and less context drift.

So stop asking "which is better." Ask "better at what." 👇
🟢 Deep, gnarly problems on a huge codebase, where correctness matters more than cost → Opus 4.7. It reasons, self-verifies, and ships a working fix.
🟡 Long autonomous workflows, web search, tool use, many steps on a budget → GPT-5.5. Reach and efficiency.
🔴 Building a product on top of these? → don't hardcode one. Abstract the model layer so you can route each job to whichever model wins it.

Here's the real lesson 👇

Models leapfrog every ~2 months now.

Loyalty to one platform isn't a strategy anymore — it's a liability.

The durable edge is knowing which model to reach for, per task, and being able to swap without a rewrite.

💡 Building agents on top of these models? Our Agentic AI Engineering course at Towards AI Academy teaches production-ready agent systems from scratch. Link in bio.

Are you just chatting with AI, or are you building with it?Most people stay at Level 1, but the real magic happens when ...
06/18/2026

Are you just chatting with AI, or are you building with it?

Most people stay at Level 1, but the real magic happens when you dive below the surface. Move from simply asking for advice to building an automated Operating Loop.

Swipe through to see the full breakdown of the AI engineering stack.

06/18/2026

🤖 GPT-5.5's benchmark scores aren't the story.

The story is that "using AI" is quietly becoming "managing AI."

The full breakdown is in the video — here's the shift that actually matters for how you work. 👇

GPT-5.5 is built around Codex, and Codex isn't a coding assistant. It's a work system.

The new way to work looks like this:

You don't do the task.

You break it into pieces, hand them to agents, and review what comes back.

In our own workflows, we run up to 20 subagents at once — some researching, some fact-checking, some attacking the draft — while we just review.

OpenAI's own finance team ran 24,000 tax forms through it and saved two full weeks.

So the skill that's suddenly valuable isn't "can you do the work."

It's three new ones 👇
🟢 Decomposition — breaking a big task into agent-sized pieces.
🟡 Delegation — writing instructions clear enough that an agent won't drift.
🔴 Verification — catching where it confidently got things wrong.

That last one is the real bottleneck now. When agents do the doing, review becomes the job.

Greg Brockman framed where this goes: ChatGPT + Codex + a browser agent merging into one delegated worker — a "super app."

The mental model to adopt today: stop thinking "tool." Start thinking "small team you manage."

💡 That exact shift — AI as a team you direct — is what our Agentic AI Engineering course at Towards AI Academy is built around. Link in bio.

06/18/2026

Transformer authors are basically the Avengers class of modern AI lol.

now has 3 of the original "attention is all you need" paper author after joined.

will gpt-5.6 > fable?

06/17/2026

The fastest AI engineers are not the best prompt writers.

Prompting matters.

But if your whole AI engineering skillset is “I write better prompts,” you are still missing the engineering part.

Here is the real skill ladder:

1. Prompt engineering: 4/10
Useful, but not enough by itself.

2. Agent frameworks: 5/10
LangChain, LangGraph, CrewAI, etc. are helpful, but frameworks change fast.

3. Tool calling: 7/10
You need to understand structured outputs, clean JSON, validation, and what happens when fields are missing.

4. RAG: 8/10
Companies do not just want chatbots. They want AI systems that work with their docs.

5. Evals: 9/10
“It looks good” is not production-ready. You need to know if the system got better or worse.

6. Backend: 10/10
APIs, databases, caching, logs, deployment, monitoring.

The AI is only one part of the product.

The person who gets hired fastest is the person who can take the model and ship the system around it.

Save this as your AI engineering skill ladder.

06/17/2026

🤔 "Claude Design vs Google Stitch — which one wins?" is the wrong question.

They're built for different moments in the same workflow.

The full comparison is in the video — here's how to actually decide. 👇

The real split isn't quality. It's cost psychology.💸 Claude Design runs on your subscription credits.

On a real project, you feel every generation — so you self-edit and explore less.🆓

Google Stitch is free (with a daily cap). That changes your behavior: you try the weird idea, test the edge case, generate five screens just to see.

Free exploration isn't just cheaper. It makes you braver.

So here's the matrix 👇

🟢 Need to explore fast, UI-only, no fear of cost → Stitch. Five screens at once, link them into a clickable prototype, even talk to the canvas with voice.
🟡 Need polish, decks, carousels, one-pagers, or websites — not just app screens → Claude Design. Stitch is UI-only; Claude Design covers way more formats.
🔴 Ready to ship → Claude Design → one-click handoff to Claude Code.

The workflow that actually works: explore freely in Stitch, then move to Claude Design to polish and hand off.

But the real headline isn't the tools.

It's that the cost of getting from "idea" to "something real" just collapsed.

The bottleneck now is taste, not tooling.📩

Want breakdowns like this first? Join the Towards AI newsletter — link in bio.

06/17/2026

congrats to Cursor for getting acquired by SpaceX

OpenAI and Anthropic should be really getting worried about this.

Cursor's composer 2.5 was trained with limited amount of compute and was so good for its price.

imagine how good their next model will be.

Their next model will be 1.5T parameter (as big as Opus and GPT) and this time not trained on top of Kimi.

Cursor is joining the SOTA model party.

Address

2810 N Church Street PMB 99625
Wilmington, DE
19802

Website

Alerts

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

Contact The Business

Send a message to Towards AI:

Shortcuts

Share

Category