Jakub Panik (๐Ÿ,๐Ÿ’ฆ) (@JakPan) header
Jakub Panik (๐Ÿ,๐Ÿ’ฆ) (@JakPan) profile picture

AI psychosis maximalist. @symbiotic_sh ex CTO @hydration_net.

Joined April 2015
Photo by Jakub Panik (๐Ÿ,๐Ÿ’ฆ) (@JakPan)Photo by Jakub Panik (๐Ÿ,๐Ÿ’ฆ) (@JakPan)Photo by Jakub Panik (๐Ÿ,๐Ÿ’ฆ) (@JakPan)Photo by Jakub Panik (๐Ÿ,๐Ÿ’ฆ) (@JakPan)Photo by Jakub Panik (๐Ÿ,๐Ÿ’ฆ) (@JakPan)Photo by Jakub Panik (๐Ÿ,๐Ÿ’ฆ) (@JakPan)Photo by Jakub Panik (๐Ÿ,๐Ÿ’ฆ) (@JakPan)Photo by Jakub Panik (๐Ÿ,๐Ÿ’ฆ) (@JakPan)Photo by Jakub Panik (๐Ÿ,๐Ÿ’ฆ) (@JakPan)
Trying the new @Google model families. Argon model in @antigravity. Same prompt compared to @grok . Grok used 0% of usage limit during 6 passes on medium and found 3/3 bugs on this PR. Argon with 2 (medium and high) used 7.5% of 5 hour limit. and 3% of weekly limit. (now showing 10% of 5 hours) Looks like display bug in agy. I stopped it before full 6 runs because I would use my limit on a medium size PR reviews... So far this seems like by far the most expensive subscription based model for the performance for PR review. So far low hopes. On the contrary, I tried Gemini Omni Flash yesterday, compared to Seedance 2.5, HY3, Kling and few others and it was BY FAR. The most efficient model of all of them for the quality you get. I would say very close second best after Seedance 2.5, probably only because it doesn't do native 1080 generation. I hope the Embedding Gemma 2 will be good. Testing soon.
2
4
440
Whaaat :O this is epic
Cloudflare's Clef beat Jev in just two weeks How: "frozen" Qwen does one prefill pass, a tiny schema head scores every answer option in parallel, with no text generated at all. That's why it's 4x faster than Jev at 2x the accuracy on some benchmarks. And it's open source! Link to HF: huggingface.co/Cloudflare/clโ€ฆ
1
240
The absolute sweet spot right now is using Opus 5.5 High as orchestrator using GPT Astra 6 High as integrator with Grok 4.7 and Astra 6 as reviewers I have tested council of Opus, GPT, Kimi, Gemini, Grok and Deepseek Grok + GPT catches 98% of issues without mistakes and loops.
1
1
4
251
3/4 Message classification Message lands into chat, model needs to understand the previous conversation semantics Is it relevant to thread? Is it answer? Is it a new topic? Is it safe to execute? Which agent should it route to? Before small but slow model. Now blink of an eye
1
208
4/4 Currently the accuracy I tested with real data is 95+% of Astra, while being 1000+x cheaper and orders of magnitude faster. Any ideas for cool classifier problem solutions? P.S. It can play doom realtime
1
1
182
2/4 Mail (ingest) classification Imagine sorting incoming mail by categories and urgency Invoices, statements, offers, leads etc... Actionable now or later? I have been using deepseek for this but jed is at least 100x cheaper and 20x times faster, less mistakes
1
265
1/4 I was just testing @typesafeai jed model on multiple classification problems. TLDR; It blows absolutely everything out of the water. It is one of the missing pieces to create responsive and cost efficient AI driven apps.
1
12
1,673
Sleep deprivation take: WoW Forever, GTA 6 and Diablo 5 are strategically timed PSYOP to slow down AI advancement.
2
16
731
I got scared for my future. More than 15 years ago, I started my programming career by being addicted to half life, warcraft 3 and then hosting unofficial WOW server. I couldn't afford the real deal. Fortunately I saw this avatar race invention. I can safely continue grinding on my 6 AI max subscriptions ๐Ÿคฃ
Choose your path as Azeroth's newest playable race: the Skyborne ๐Ÿชฝ ๐Ÿ”ฎ Travel the Arcane road with the Alliance OR ๐Ÿ‚ Master the Elements with the Horde
1
1
5
1,109
Jakub Panik (๐Ÿ,๐Ÿ’ฆ) retweeted
Formal verification of smart contracts is the last line of defense. That is where we make our stand. Else, get ready to be rugged.
kimi k3 also did a zero-click arbitrary command execution in Telegram Desktop and iOS app (aslr pinned, one gadget away from full rce).
1
1
8
1,575
I almost shipped competing product to @ether_fi at one point in my life ๐Ÿฅฒ It's hard to compete tho as it's almost perfect. I have used 10+ crypto cards and none of them offer anything close. All of the bullshitters farming twitter affil never used the it. People need to give hats off to @MikeSilagadze and the team because self custodial protocol built on top of decentralized-ish vaults offering borrowing against collateral for payments doesn't exist anywhere else in this scale, fee ranges and transparency. Their behavior during crises was impeccable. Not enough props were given. Yeah you can get lounge access, but you can't get 3% ETH cashback while borrowing against your assets while you still own them. Thus no affiliation link is given. Use the one to get 10% off claude or whatever. Just get the superior product.
7
1
30
4,950
Jakub Panik (๐Ÿ,๐Ÿ’ฆ) retweeted
Replying to @AnthropicAI
Dev-motional loop with @AnthropicAI New model Maxxing - no sleep It deletes 3 hours of work and 100$ of API credit runs at 3AM Harden till 6AM to run while you sleep Wake up to zero progress and nerfed model "Sorry, my mistake" Downgrade to 20$ and switch back to @OpenAI
1
3
2,608
Jakub Panik (๐Ÿ,๐Ÿ’ฆ) retweeted
Replying to @blader
31
15
27
227
465,161
It rips but I guess I need more subscriptions. Few minutes into the usage reset ๐Ÿ˜„
1
1
423
Fable feels at least +30 IQ. Can't go back @AnthropicAI
2
4
317
Also it's 5AM again ๐Ÿฅน
178
Deepseek 4 is legit hack for benchmarking and fine-tuning. You get < 3s response with deepseek-4-pro at medium reasoning. 20+ parallel sessions.
1
2
431
compared to any other frontier model at 7s+ per response in this case. In Asia, with similar results and the cost being so cheap it means rapid iteration and significant improvement.
1
1
294