Veros

🇮🇩 VP Product & Solutions - | 3x K8s | 2x MikroTik | Open Source Software | Building @review_codes

[Veros @boldnames]

Logika membuka jalan kebenaran.

Kalkulasi mentah terhadap entropi (kekacauan) menuju ekuilibrium (kestabilan), menjadi bukti bahwa tidak ada rumus paling efisien kecuali Tauhid dalam vakum realitas.

[Veros @boldnames]

@papanberjalan @Canonical Ini MicroCeph nya MicroCloud, standalone atau seperti apa ya, versinya boleh di sharing mas?

[Veros @boldnames]

@S1r1u5_ Count me in pls

[Veros @boldnames]

@mohammed_hijab Agreed.

[Veros @boldnames]

@Jason I'm in, building @review_codes

[Veros @boldnames]

@GergelyOrosz If they're doing this and SLA still sucks it won't be good for them.

[Veros @boldnames]

Passed 5 million Lines of Codes of OSS comprehensive cybersecurity audit (OWASP, CWE, OSV, 0-days, etc) on @review_codes alpha stage.

If there's any OSS worth auditing, do let me know please we will audit that for free (End-to-end AI Scanner does cost a lot).

[Veros @boldnames]

@yasser_elsaid_ Agreed.

https://t.co/bp7LJ1oxDh

[Veros @boldnames]

If you are in AI wrapper run.

Pivot to Harness engineering.

Nobody knows jack shit there still.

[Veros @boldnames]

Are there any @GroqInc like inference providers, but instead of in/out cost, they are rated by API Call instead?

[Veros @boldnames]

@MoiTrades @0xSero Yes i think so.

[Veros @boldnames]

Introducing Mythos alternative to scan your entire codebase for cybersecurity threats.

👉https://t.co/LcLcBuEjMM

We coupled deterministic algorithm with agentic workflow that resulted to a much more powerful, context-aware code review workflow than just a regular pattern-based like ASTs, or Taints scanner.

With current *Alpha* stage, we now achieved a stable 70-95% accuracy throughout our internal scan testing.

[Veros @boldnames]

@manicode I'd argue hybrid models work best deterministic (i.e SAST) combined with LLM intelligence, this is what I am building with https://t.co/NneEzuImjE.

[Veros @boldnames]

@zackkanter Kubernetes is not a Scam, just mostly misunderstood.

[Veros @boldnames]

@chrisgpt ".... is weeks away from .... " energy.

[Veros @boldnames]

@0xSero Thank you appreciate the pointers. Where I live the 6000s cost 30-50% more...

[Veros @boldnames]

@0xSero Nice, what is your setup? any pointers for this please?

[Veros @boldnames]

@thsottiaux When GPT5.5?

[Veros @boldnames]

@OpenAI I thought GPT 5.5 was promised to us?

[Veros @boldnames]

Claude Opus 4.7 release announcement just proved this post.

https://t.co/4vTBkmTfBA

[Claude @claudeai]

Introducing Claude Opus 4.7, our most capable Opus model yet.

It handles long-running tasks with more rigor, follows instructions more precisely, and verifies its own outputs before reporting back.

You can hand off your hardest work with less supervision.

[Veros @boldnames]

You can save yourself a lot of time when co-working with AI when you factor this at all times (with practical remedy towards the end of the post):

> AI output is NEVER a constant 100% "accurate" at anything.

There is only 1 fundamental and technical Truth inside the mind of AI:

> Calculative or analytic truth, where LLM do calculations based on data it's trained on towards an output based on user input (minimizing entropy).

In simpler terms, basically an LLM is an expensive semantic calculator, instead of a conventional calculator which processes deterministic numbers.

You can input to a SOTA model an 'objective' or deterministic Truth like "what is 1+1?" and I'm sure with 99.99% probability (on a fresh memory session) it will always output the expected result.

But the output "calculation" shifts, when you have an existing memory (e.g memory pollutions, context memory overload, or you ask the AI to answer wrongfully).

And this goes without saying, the output calculation dramatically shifts when you input to a SOTA model a non-objectives (e.g how is your day? what do you think of x? etc), it will never respond uniformly every time (unless again, you have existing memory in the session, or you tell it to).

In short, ALWAYS treat AI output as "computed approximation".

And then you may wonder wtf does this all means to my daily AI usage?

Here's for your productivity boost if you care about accuracy or precision:

> Mathematically speaking, give the AI an input or ask it to reverify its own output at least once for a mathematically better output approximations.

> And before that, always make sure your memory context usage within a session is in the "healthy" range (that's for another post).

If you like this kind of content where a human still write stuff, then drop a follow.

[Veros @boldnames]

Claude Opus 4.7 release just proved these tips.

Wonder what other tips of mine that's also right, follow for more ;)

[Veros @boldnames]

You can save yourself a lot of time when co-working with AI when you factor this at all times (with practical remedy towards the end of the post):

> AI output is NEVER a constant 100% "accurate" at anything.

There is only 1 fundamental and technical Truth inside the mind of AI:

> Calculative or analytic truth, where LLM do calculations based on data it's trained on towards an output based on user input (minimizing entropy).

In simpler terms, basically an LLM is an expensive semantic calculator, instead of a conventional calculator which processes deterministic numbers.

You can input to a SOTA model an 'objective' or deterministic Truth like "what is 1+1?" and I'm sure with 99.99% probability (on a fresh memory session) it will always output the expected result.

But the output "calculation" shifts, when you have an existing memory (e.g memory pollutions, context memory overload, or you ask the AI to answer wrongfully).

And this goes without saying, the output calculation dramatically shifts when you input to a SOTA model a non-objectives (e.g how is your day? what do you think of x? etc), it will never respond uniformly every time (unless again, you have existing memory in the session, or you tell it to).

In short, ALWAYS treat AI output as "computed approximation".

And then you may wonder wtf does this all means to my daily AI usage?

Here's for your productivity boost if you care about accuracy or precision:

> Mathematically speaking, give the AI an input or ask it to reverify its own output at least once for a mathematically better output approximations.

> And before that, always make sure your memory context usage within a session is in the "healthy" range (that's for another post).

If you like this kind of content where a human still write stuff, then drop a follow.