iiinova — where your digital self keeps evolving

ai•Lab

@ailab.ee

Building practical AI solutions for complex problems. No hype, just engineering. We build custom applications, autonomous agents, and provide real-world AI strategy.

Posts Replies Media

ai•Lab @ailab.ee · 28/07/2026

The models you choose make a big difference. Do your homework before making any long-term commitments.

AI Lab infographic comparing LLM API prices in USD per 1 million tokens, shown as input/output, with Artificial Analysis Intelligence Index v4.1 scores. Rose represents proprietary models; teal represents open-weight models. Proprietary: GPT-5.6 sol—59, $10/$45; terra—55, $5/$22.50; luna—51, $2/$9; Claude Fable 5—60, $10/$50; Claude Opus 5—61, $5/$25. Open-weight: Kimi K3 via Moonshot AI—57, $3/$15; GLM 5.2 via Z.ai—51, $1.40/$4.40; StreamLake—51, $0.7525/$2.365; AkashML—51, $0.77/$2.42. Main findings: median open-weight output price is $3.41 versus $25 for proprietary models, 7.3 times lower. At index 51, StreamLake costs $2.365 output versus $9 for GPT-5.6 luna, 3.8 times lower.

020

ai•Lab @ailab.ee · 24/07/2026

Don’t just blindly automate your company’s work with whatever LLM you’re used to. This graph is a perfect illustration of why it’s important to pick the right model for every task. The cost difference over the years will be huge.

Bar chart titled “Cost per Task,” showing weighted average cost in USD per Intelligence Index task; lower is better. Costs rise from DeepSeek V4 Pro at $0.04 and GPT-oss-120b at $0.06 to MiniMax M3 at $0.12, Nemotron 3 Ultra at $0.24, Muse Spark 1.1 at $0.26, Grok 4.5 at $0.31, GLM-5.2 at $0.32, Gemini 3.6 Flash at $0.50, Kimi K3 at $0.95, GPT-5.6 at $1.04, and Claude Fable 5 with fallback at $2.75.

020

ai•Lab @ailab.ee · 08/06/2026

Yes! This 👇

X.com post by Peter Steinberger (@steipete): “Here’s your monthly reminder that you shouldn’t be prompting coding agents anymore. You should be designing loops that prompt your agents.” Timestamp: 21:58 · 07/06/2026 · 4.3M Views. Engagement: 1.3K replies, 1.9K reposts, 16K likes, 12K bookmarks.

010

ai•Lab @ailab.ee · 18/03/2026

Together with @heysec.com we built a fully autonomous cybersecurity research factory. It researches prompt injection vulnerabilities, builds research cases with PoC exploits, creates firewall rules to mitigate these threats, and iteratively tests those against active exploits.

Dashboard screenshot with status bar: “Firewall,” “Agent Engine,” “8 agents,” “0 pipelines,” and button “START PIPELINE.” Main panel “RESEARCH & TRIAGE” shows 111 Sources, 794 Entries, 231 Threats, 8 Engines, 15 Engine Limited; status line: 2 Awaiting, 0 Queued, 0 Triaging, 225 Promoted, 0 Dismissed, 5 New (24h). “AGENTS (24H)” lists: Researchers 0 Pending, 0 Running, 0 Done, 0 Failed, 0 Timeout; Discoverers 0 Pending, 0 Running, 4 Done, 0 Failed, 0 Timeout; Triagers 0 Pending, 5 Running, 76 Done, 135 Failed, 0 Timeout; Pipeline 0 Pending, 0 Analysis, 0 Rules, 0 Testing, 12 Done, 3 Failed, 0 Cancelled. “NEW ITEMS ADDED” table: Sources 24H 2, 7 DAYS 11, 30 DAYS 71; Researches 5, 32, 417; Threats 34, 35, 231; Rules 29, 30, 262. Right-side panels: “PIPELINES” Active 0, Completed 59, Total 74; “RULES” Draft 0, Testing 5, Approved 206, Published 0, Total 262; “LEARNER” Completed (7d) 3, Findings (7d) 15, Changes (7d) 10, Total runs 37.

010

ai•Lab @ailab.ee · 12/03/2026

These agents' confusion moments are a great example of this other 20-30%. 👉 ailab.ee/we-turned-openclaw-into-a-…

iMessage conversation with Andy. Andy: “Looks like food wasn't logged today — hard to track the deficit without it. Even a rough log helps.” (2m) Reply: “It was. But I guess the sync hadn’t run yet. Is the data there now?” (Now) Andy: “Could you remind me what you're referring to? I want to make sure I check the right data — the last few topics were moving quotes, health stats, and earlier the morning brief. Which sync are you asking about?” (Now)

011

ai•Lab @ailab.ee · 11/03/2026

We’re putting the OpenClaw EA to the real test and asking it to handle the move of our offices. Let’s see how that goes 🤞 👉 We published more about what this AI EA really is all about: ailab.ee/we-turned-openclaw-into-a-…

Chat screenshot message: “Sending to 10 more providers now. Also — Moving24 already replied while I was working on this: • €85/h + VAT — 2 movers + truck • Free packing materials included • Available. So you’ve got one quote in already. I’ll flag the others as they come in.”

010

ai•Lab @ailab.ee · 10/03/2026

You can now also listen to all our blog posts 🎧 Give it a try on the latest blog post about how we built an EA on OpenClaw: ailab.ee/we-turned-openclaw-into-a-…

Audio player controls (play button, “0:03 | 11:00,” volume icon, star button, double-down arrow button, and “1x”) above text that reads: “If you haven’t come across OpenClaw yet, here’s the short version: it’s an open-source personal AI assistant that runs on your own machine. Mac, Windows, or Linux. You talk to it through whatever chat app you already use”

110

ai•Lab @ailab.ee · 10/01/2026

Version 2 of the Support Agent we built a while back is coming along really well.

Support Agent web app screenshot showing “Flow Builder: Support Ticket Flow.” Left sidebar menu: Dashboard, Runs, Flows (selected), Agents, Tools, Prompts, Plugins, Users, Settings. Step Types panel lists: Filter (“Filter or validate input data”), Agent (“Execute an AI agent”), Decision (“Branch based on conditions”), Validation (“Validate output with LLM”), Action (“Execute external action”), Transform (“Transform data format”). Main canvas flow: “Classify Ticket” (Agent, Entry) → “Check Response Needed” (Decision) branching to “Research & Process” (Agent, highlighted) or “Close Ticket” (Action); “Research & Process” → “Check Escalation” (Decision) branching to “Escalate Ticket” (Action) or “Validate Response” (Validation) → “Compose Final Response” (Agent). Right panel shows “Research & Process” settings: Name “Research & Process,” Entry Point toggle, Select Agent “Test Support Agent,” link “View/Edit Agent Configuration,” Connections with Incoming (from) “Check Response Needed” and

020

ai•Lab @ailab.ee · 06/01/2026

It doesn’t always have to be a large and complicated application to improve the user experience with AI. Sometimes just a tiny addition will add the most value. Like this integrated spellchecker using Gemini 3 Flash.

Social media post composer screen with buttons “Cancel” and “Post” and share icons; text box labeled “Post 1” with counter “162 / 280” contains: “It doesn’t always have to be a large and complicated application to improve the user experience with AI. Sometimes just a tiny addition will add the most value.” A “Proofread” button is highlighted, with a green arrow pointing to it.

011

ai•Lab @ailab.ee · 24/12/2025

Happy holidays! 🎄

011

ai•Lab @ailab.ee · 20/12/2025

Here’s how we built our AI-powered customer support agent: ailab.ee/projects/intelligent-custo…

011

ai•Lab @ailab.ee · 21/09/2025

Got the sudden urge to buy this entire look? (That hat though!) That's the magic of styled outfits working on you. This image was AI-generated in minutes. 👉 Here's how we did it: ailab.ee/2025/09/how-...

020

ai•Lab @ailab.ee · 27/08/2025

AI image generation isn't just for cool edits. Smart companies use it for product variations, marketing visuals, virtual try-ons, and turning blueprints into renderings. The question isn't if AI will change your business – it's how fast you'll adapt.

000