Shmulik Cohen

AI Engineer at Apiiro | student at Tel Aviv | I run “AI Superhero” | In pursuit of truth and goodness

[Shmulik Cohen @shmulc8]

Relevant papers:
https://t.co/hZmyslpA04
https://t.co/lFgToWbrA0
https://t.co/jhmONEd1jC

[Shmulik Cohen @shmulc8]

We broke 3 published math bounds for hypercube metric dimension β(Qₙ):

• β(Q₂₄) ≤ 13 (was 14) — 14 test attempts
• β(Q₂₆) ≤ 14 (was 15) — 15 test attempts
• β(Q₂₉) ≤ 15 (was 16) — 16 test attempts

Here's how we did it with AI & sign matrices 🧵👇

[Shmulik Cohen @shmulc8]

I built an interactive site to visualize the deduction: pick any answer sheet (from 1.07B possibilities!) and watch the algorithm collapse it to 1 in 17 attempts live: https://t.co/1xwNgG8RQ4

[Shmulik Cohen @shmulc8]

Using Anthropic's Fable 5, we searched millions of sign matrices and improved the bounds.

Don't trust the AI blindly: each record is certified by a small matrix verified in seconds by 2 independent verifiers: https://t.co/VMkb4xJcbs

[Shmulik Cohen @shmulc8]

This is Erdős & Rényi's 1963 coin-weighing problem in disguise.

Surprise: exact values of β(Qₙ) stop at n=13! Beyond that, only upper bounds exist.

[Shmulik Cohen @shmulc8]

It started with a riddle: You take a 30-question true/false test, but after each attempt you only learn your total score. How many attempts guarantee 100%?

(The answer is 17, known since 1966)

[Shmulik Cohen @shmulc8]

@The_MRI תנסה את agy ותראה מה זה מהירות

[Shmulik Cohen @shmulc8]

@Hacker0x01 Just finished it, very HARD!

[Shmulik Cohen @shmulc8]

I'm playing with an idea: can LLMs use Alloy's software-modeling capabilities to find real protocol bugs?

GPT-5.6 Codex + Alloy found one:
issue parent → revoke parent → delegate again → fresh child accepted

Spec → Alloy → real Go PoC. One missing check caused it.
@OpenAI

[Shmulik Cohen @shmulc8]

Full writeup, method, and every audited disagreement 👇
https://t.co/GTi5FrgYh6
Ran on @novita_labs. Hy3 was listed at $0 when I ran it.

[Shmulik Cohen @shmulc8]

Tencent Hy3 vs GLM-5.2 on a 165-alert SASTBench queue.

I checked every model and benchmark disagreement in source.

Hy3: 14/19 real, 5 false alarms.
GLM: 13/19, 7 false alarms.
Full writeup below 👇

@TencentHunyuan @novita_labs #Hy3Novita

[Shmulik Cohen @shmulc8]

Full writeup: what worked, what hallucinated Chinese gibberish, and the "forgetting curve" idea that made me rethink context windows:
https://t.co/Y0qj4j5TK7
Try DeepSeek-OCR 2 yourself, my signup link: https://t.co/LbfOD6Acr1

[Shmulik Cohen @shmulc8]

I ran 48 pages of my most cursed thesis PDFs through DeepSeek-OCR. Total cost: less than half a cent.

But the price isn't the wild part. It's the theory underneath: that text might be a terrible way to store text.

Full breakdown, run on @novita_labs 👇

[Shmulik Cohen @shmulc8]

@GutmanYotam מעניין, לאחרונה כתבתי גם פוסט בסגנון https://t.co/PMAY37AVpw

[Shmulik Cohen @shmulc8]

Full build guide:
https://t.co/kUXpx3YkUs
I dry-ran the whole stack on Novita's serverless endpoints before cutting the cord. Try the same open models with my signup link: https://t.co/XAgW466GCw

[Shmulik Cohen @shmulc8]

Your spaceship just lost contact with Earth. No ChatGPT, no PyPI, no StackOverflow. Your astronauts still need to ship code.

I built the full offline AI coding rig with @novita_labs as the launchpad: Qwen3-Coder-480B, local doc search, offline StackOverflow.

Build guide 👇

[Shmulik Cohen @shmulc8]

@david_lisovtsev פטור מתור מסכים שדיי מיותר, תו נכה כדאי מאוד שיהיה אבל התנאים צריכים להיות הרבה יותר נוקשים

[Shmulik Cohen @shmulc8]

@Idanhcadash ממליץ על הרצאות של יושי פרג'ון בנושא, הוא מתעסק בזה ולא מעט וממש כיף לשמוע אותו

[Shmulik Cohen @shmulc8]

@GadiAleks זה נורא, אפשר להגיע מהקישור הזה
https://t.co/ygXAXK4RTQ

[Shmulik Cohen @shmulc8]

@FeedTechILUncen זוכרים מה קרה עם אינטל כשהם קצת הורידו תנאים?