Kaluxik

Programming, automating, occasionally investing. Production AI - RAG, multi-agents, automation pipelines. Shipped, not demoed. Warsaw 🇵🇱

[Kaluxik @kaluxik]

@svpino Code is no longer the product, it is just raw material. AI made writing code cheap, so the hard part is now deciding what to build, testing it, and keeping it secure. Pumping out more software is useless if you are building the wrong things.

[Kaluxik @kaluxik]

@Jeande_d I saw this on multi-step extraction. Adding reasoning tokens tripled latency and doubled cost, but error rates on edge cases barely moved. The model just spent 800 tokens rewriting the same flawed assumption.

[Kaluxik @kaluxik]

@svpino Running four agents in parallel is great until you have to review four diffs. What we actually need next is an evaluator agent to run tests, compare state changes, and pick the best implementation automatically.

[Kaluxik @kaluxik]

@emollick The bottleneck on irregular tasks is verification. If an agent handles a dispute with even a 5% error rate on policy codes, you still spend 15 minutes checking raw PDFs. High human-in-the-loop overhead kills the ROI.

[Kaluxik @kaluxik]

@SIGKITTEN We didn't get smarter. We just automated the effort of pretending we were curious.

[Kaluxik @kaluxik]

@OfficialLoganK In multi-agent pipelines, offline evals are barely 5-10% of total spend. The real burn is online LLM-as-a-judge for continuous monitoring and guardrails on live traffic, which easily eats 30% of the token budget.

[Boris Cherny @bcherny]

We've been working on this with customers for a while. Mythos-class models require additional safety measures and enterprises need to meet their own privacy and compliance rules. Customers can own and control their own data and Anthropic retains none. It’s coming this fall.

[Kaluxik @kaluxik]

@nickbaumann_ Strongly agree. Most people have no idea what's already possible. For them AI is still just a chat box.

[Kaluxik @kaluxik]

@Im_IrushiK Odds are it won't be an AI company at all, just a lab renting the model like everyone else. Whoever builds it gets paid for tokens, whoever uses it gets the Nobel.

[Kaluxik @kaluxik]

The gap here is unreal. $7,400/employee vs $12.

The top 1% aren't using a better model. They stopped chatting with it and gave it part of the work.

The difference isn't the AI. It's what the company did with it.

[Olivia Moore @omooretweets]

The median company is spending $12 / employee / month on AI

The top 1% are spending $7,500 / employee / month

Not sure we've ever seen an adoption gap quite like this

(h/t @tryramp data, @a16z)

[Kaluxik @kaluxik]

@YourAnonOne It’s all because they added Gemini to every product they own. Curious how many of them use it only in chat.

[Kaluxik @kaluxik]

@burkeholland This is the biggest problem with local models - TPS and no parallelism.

[Kaluxik @kaluxik]

The problem with local models won’t be answer quality. It’ll be tokens/sec.
Chat forgives 18 tokens/sec, but agentic loop doesn’t.

[Burke Holland @burkeholland]

Been using Qwen 3.8 27B (Q4) locally on 64GB of VRAM.

Here is the verdict: SLOW

18 tps with ZERO system prompt to process and that degrades significantly with a harness system prompt and as the context window grows.

RIP if you have to compact.

I had it implement this PRD https://t.co/y0M0pDv61x and it's been running for 6 hours.

By comparison Grok 4.6 and Kimi K3 hosted finished in about ~30 minutes.

High hopes, but these 27B variants are too dense. This is not a consumer grade local model - and I consider consumer grade to be anything up to $5000.

[Kaluxik @kaluxik]

@0x0SojalSec Actually, that could be his face after reading the news.

[Kaluxik @kaluxik]

@TheAhmadOsman Fable 5 on single RTX 5090 in next 6 months?

[Kaluxik @kaluxik]

Company that can’t make images is in talks to buy one that makes video in real time.
Officially it’s about chip efficiency. But maybe text-only stopped looking like enough?

[Bloomberg @business]

Anthropic is in talks to buy the artificial intelligence startup Decart for about $6 billion, sources say https://t.co/2P60RI4B6k

[Kaluxik @kaluxik]

@birdabo I think we are around 6 months to 1 year away from an answer.

[Kaluxik @kaluxik]

Opus 4.6 shipped Feb 5, 2026. It held the crown for exactly 10 weeks.

Six months later, a 27B model that keeps pace with it on parts of the benchmark suite runs on 17GB of RAM - on own desktop.

If your architecture assumes today's frontier is next year's frontier, it doesn't.

[Qwen @Alibaba_Qwen]

27B on 17GB RAM. Are you ready to create something incredible? 😎
Thanks for highlighting it! @UnslothAI

[Kaluxik @kaluxik]

@TheAhmadOsman Genuinely curious when this starts hurting the API providers. At what point does "good enough on my own GPU" beat "slightly better behind a paywall" for most production workloads?