Parik

ai builder | documenting the journey | 0 → ?

[Parik @pparik__]

Starting from 0
No audience
No connections
No shortcuts
Just curiosity, consistency, and a lot of AI

I'll share everything I learn, build, and discover

First goal: 100 followers
Let's see how far this goes

[Parik @pparik__]

@TheAhmadOsman If you look into the future, then yes, everything is correct

[Parik @pparik__]

@higgsfield_ai Both models are good, no one is better, no one is worse, they just do it in their own way

[Parik @pparik__]

@CryptoKaleo Wasn't there already a bullish season on Robinhood a couple of days ago, now it's been quiet for a couple of days

[Parik @pparik__]

@designcoursecom Perhaps, during the process of work, Fable seems more human, since it can ask more questions, while GPT 5.6 is more like a mechanism, it will precisely do what it says and that's it, nothing more

[Parik @pparik__]

Hugging Face just disclosed an autonomous-agent intrusion. The forensics detail is the real post

Initial access: malicious dataset → remote-code loader + config template injection on a processing worker. Then node access, credential harvest, lateral move across clusters. Swarm of short-lived sandboxes. >17k attacker events in the log.

They say no evidence public models / Spaces / supply-chain packages were tampered. Partner impact still under assessment. Rotate tokens anyway.

Part that should stick: commercial API guardrails blocked submitting real attack payloads for analysis. HF ran forensics on open GLM 5.2 on-prem so exploit data and credentials never left their network.

Attacker had no usage policy. Defender's first-choice hosted models refused the evidence.

Dataset loaders are production attack surface. If you process untrusted data with code paths, assume agent-speed offense is already cheaper than your weekend on-call

[Parik @pparik__]

"Open 3T-class" is marketing until the checkpoint is on your disk

Moonshot shipped Kimi K3 Jul 16: 2.8T MoE, 1M context, native vision, API live. Full weights promised by Jul 27.

Independent numbers that matter today:
- AA Intelligence Index 57 → ~#4 of 189 (behind Fable 5 / GPT-5.6 Sol max; ~Opus 4.8 / GPT-5.5 class)
- Arena Frontend Code #1 at 1679, 76% pairwise win rate, #1 in 6/7 domains
- API $3 / $15 per 1M (cache hit $0.30). ~62 tok/s, ~1.99s TTFT on first-party
- AA still lists it as proprietary until weights land

Param count is a research artifact. A laptop does not run 2.8T.

The real story is Frontend Arena #1 + AA 57 + price/task vs Fable/Sol - not the trillion-parameter press release.

Wait for the download, the third-party harnesses, and real agent sessions after Jul 27.

[Kimi.ai @Kimi_Moonshot]

Introducing Kimi K3: Open Frontier Intelligence

🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal
🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts
🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost
🔹 Built for long-horizon agentic coding and self-evolving workflows

Kimi K3 is now live on on https://t.co/zrk6zZxZUo, Kimi Work, Kimi Code, and the Kimi API.
Open Weights by July 27, 2026.

🔗 API: https://t.co/XCrgjXAqMw
🔗 Tech blog: https://t.co/YTfiMSNM1f

[Parik @pparik__]

@CodeByNZ This is the first time I've seen this. Kimi has such a respectful attitude towards her users. I like it.

[Parik @pparik__]

@theo I also heard that many people praise it, but I also heard criticism about the number of tokens, which is quickly spent at the moment

[Parik @pparik__]

@alex_prompter Isn't it too early to shift everything? Do you need more days for testing?

[Parik @pparik__]

@bridgemindai in terms of the Kimi code itself, is it enough for you, does it cover your needs?

[Parik @pparik__]

@haider1 Don't forget about the errors that gpt 5.6 had, and how it deleted all the files for some users

[Parik @pparik__]

@bridgebench Yes, but it ranks 9th in the code ranking

[Parik @pparik__]

@mark_k @elonmusk It was high time to do this, now we have even more desire to use Premium+

[Parik @pparik__]

@bridgemindai I saw almost exactly the same game when grok 4.5 came out, but maybe it's a slightly different game.

[Parik @pparik__]

@bridgemindai I expected cheaper prices, with such prices, they should not only be on par with Opus 4.8, but should be much stronger, on par with Fable and GPT 5.6, only in this case people will use it

[Parik @pparik__]

@mweinbach But ideally, a couple of days are still needed for testing, so that people can test using different methods

[Parik @pparik__]

@Jtyles A good use of an extra hour, a good habit

[Parik @pparik__]

@XFreeze The open source nature and power of grok makes it a strong competitor in the AI ​​race

[Parik @pparik__]

@bridgemindai But it's not available to everyone yet, we still have to wait