What we delivered wasn't "another bot", it was a system that reasons.
At HumAIns we teach agents how to think for each use case while staying aligned with enterprise values.
This project proved it: intelligence can be configured, not handcrafted.
On to the next impossible.
Three weeks. Three complex AI agents. One impossible deadline.
I'm incredibly proud of what our team just pulled off for IEC - Israel Electric Corporation @IecIsrael ืืืฉืื ืืืฉืจืื
When I first realized what we'd agreed to, I was in shock:
โ Three conversational proactive autonomous AI agents
โ Real-time avatars with dedicated URLs
โ Live system integrations - orders and SMS confirmations
โ Real-time Hebrew speech with dynamic Hebrew gender adaptation
โ Super-complex energy sector use cases
๐๐ฒ๐น๐ถ๐๐ฒ๐ฟ๐ ๐๐ถ๐บ๐ฒ: ๐ง๐๐ฅ๐๐ ๐ช๐๐๐๐ฆ.
To put this into perspective: Our competitors need 8 weeks to learn a SINGLE use case before running a proof of concept...
We delivered three production-ready agents in less time than it took to start testing one.
How?
Because https://t.co/1H0bujIzl0 isn't a chatbot builder nor an LLM API wrapper.
It's a cognitive framework with: โ Advanced memory architecture that remembers context across conversations โ Goal-oriented reasoning that adapts strategies in real-time โ Persona capabilities that match communication styles โ Hebrew linguistic AI that dynamically adjusts gender, formality, and cultural tone.
We're the ๐ผ๐ป๐น๐ company that's solved it.
Hebrew isn't just grammatically complex - it's culturally nuanced. Every verb, adjective, and sentence structure changes based on who you're speaking with. Male or female. Formal or casual. Young or older.
Most AI agents sound robotic or worse - culturally tone-deaf - in Hebrew.
Ours sounds like they actually understand who they're talking to.
But here's what really matters:
This wasn't just about speed. It was about proving something fundamental:
๐ช๐ต๐ฒ๐ป ๐๐ผ๐ ๐ต๐ฎ๐๐ฒ ๐ฎ ๐ฐ๐ผ๐ด๐ป๐ถ๐๐ถ๐๐ฒ ๐ข๐ฆ, ๐๐ผ๐ ๐ฑ๐ผ๐ป'๐ ๐ฏ๐๐ถ๐น๐ฑ ๐ฎ๐ด๐ฒ๐ป๐๐ ๐ณ๐ฟ๐ผ๐บ ๐๐ฐ๐ฟ๐ฎ๐๐ฐ๐ต. ๐ฌ๐ผ๐ ๐ฐ๐ผ๐ป๐ณ๐ถ๐ด๐๐ฟ๐ฒ ๐ถ๐ป๐๐ฒ๐น๐น๐ถ๐ด๐ฒ๐ป๐ฐ๐ฒ.
That changes everything.
Where others spend months custom-coding each use case, we're configuring cognitive capabilities that already exist.
The next milestone?
Enabling clients to test-drive completely new use cases in hours.
Eventually, minutes.
Because when AI is brilliant - not just automated - it's: โ Fast to deploy โ Safe to trust โ Reliable at scale
Massive thanks to the Humains team - to @AvishaiShraga who owned this project and even took a flight to Eilat to ensure it works perfectly, and to Dan Bystritsky Raz Kronenberg and Ben-Etzion Yaron who made the impossible routine, and to IEC for believing we could deliver what seemed impossible.
This is just the beginning.
#AI #HumAIns #Innovation #IEC #HebrewAI #CognitiveOS
Teach the model how to reason for your specific use case. You can achieve great results with very small models. You don't need a full and coherent chain of thought, key points will suffice.
A small note on Paul's tweet, you can also gain much improvement with just prompting.
Big claim in this paper.
A 3.8B model can match GPT-4o on documentโgrounded answers using a simple reasoning step.
a short check plus light fineโtuning beats scale.
On the FACTS Grounding test they are within 5 points, and the small model is about 19x cheaper.
The task checks if answers come only from the provided document, with no guesses.
Their method adds a brief check before writing, lists the claims, compares them to the context, and keeps only supported parts.
Small models need fineโtuning to follow that protocol, and the combo gives the big jump.
Fineโtuning teaches process discipline, not new facts, so answers without the check stay flat.
The check adds about 3% to 5% tokens and avoids costly thinking modes while keeping outputs steady.
It also reduces answer variability, so behavior is more predictable.
----
Paper โ arxiv. org/abs/2510.25933
Paper Title: "Humains-Junior: A 3.8B Language Model Achieving GPT-4o-Level Factual Accuracy by Directed Exoskeleton Reasoning"
Big claim in this paper.
A 3.8B model can match GPT-4o on documentโgrounded answers using a simple reasoning step.
a short check plus light fineโtuning beats scale.
On the FACTS Grounding test they are within 5 points, and the small model is about 19x cheaper.
The task checks if answers come only from the provided document, with no guesses.
Their method adds a brief check before writing, lists the claims, compares them to the context, and keeps only supported parts.
Small models need fineโtuning to follow that protocol, and the combo gives the big jump.
Fineโtuning teaches process discipline, not new facts, so answers without the check stay flat.
The check adds about 3% to 5% tokens and avoids costly thinking modes while keeping outputs steady.
It also reduces answer variability, so behavior is more predictable.
----
Paper โ arxiv. org/abs/2510.25933
Paper Title: "Humains-Junior: A 3.8B Language Model Achieving GPT-4o-Level Factual Accuracy by Directed Exoskeleton Reasoning"
People describe Israel as having a conflict with Iran or going to war as if we have a conflict over a border or territory or land.
But the only real conflict we have with Iran is that their governing regime says โDeath to Israelโ and we refuse to comply.
I think it's exactly how The Matrix arc begins?
Two AI agents on a phone call realize theyโre both AI and switch to a superior audio signal.
แตแตแตแตหกแตแตแตสณหข แตหขแตแต แตแตสทแตแตแต หกแถฆแตสณแตสณสธ แตแต แตแตแตแต แตสฐแต แดฌแดตหข แถแตแตแตแตโฟแถฆแถแตแตแต แถ แตหขแตแตสณ
๐งตWith 500K+ views this fake list of "Zionist arguments" requires a full debunking. Every point Shehada makes is a lie, misrepresentation, strawman, and fake history (lots of it). Detailed thread below: 1/
BEAST GAME EPISODE 3 IS OUT NOW!
To celebrate, I'm giving away $100,000 total to 10 random people who like and retweet this post!
Go watch it here: https://t.co/Yntf9E7FTN