Andrey Kolesnikov

@asaio87 I’m one of those idiots I guess. 1M MAU app. Most of our growth and feature innovation came after AI usage spread throughout the company. Exited too, made investors happy. All AI.

Only idiots think AI has changed everything when it comes to building successful apps.

@DanielSmidstrup They are capacity constrained. The rest of the world is fine using Nvidia, for Google it’s declaring an L in GPU race. Hence the focus on TPU8

Seriously, how can Google still not be leading the AI race with all it's resources

@steipete You’ll like the weather.

Finally got my visa sorted out and moving to San Francisco, just in time for MS Build and OpenClaw’s after hours! luma.com/OpenClaw-GitHub

@ChrisJBakke 5x rev is very rich for SMB.

For the last 6 years I’ve been buying well-run small businesses for 5x earnings. In the first 30 days, I take the websites offline, move the companies to sad office parks with drop ceilings, install fax machines at the front desk, and bringing in 75 year old actors to pose as the CEO. I then sell the companies to people with MBAs for 10x revenue so that they can feel useful “turning the company around”

@TheChiefNerd BG is spot on with his Frankenstein take. Loudest case for local AI - freedom of intelligence.

🚨 BILL GURLEY: “I would encourage people to read as much as they can about Anthropic … I don't think they think they're writing software. I think they're midwifing a deity.” JASON: “I know some of these folks … They believe they're so powerful, that they can create God.”

@antirez How can you forget when reminders are everywhere. I’m keep cancelling subscriptions, banks and other services that haven’t evolved. Some are still running JSX, ASP and other ancient frameworks.

Many of you forgot too fast the insane amount of shitty software we had to see and suffer in the pre-AI era.

@omooretweets Just like Sam and Dario in India.

Self-driving cars are fun because you never see competing SaaS products having a literal standoff in the street

@witcheer @NousResearch 27b is not really designed for multi-step and context recall. Something bigger needs to feed it isolated chunks of bound context, it rips.

Which local LLM best drives an agent? I built a benchmark for pairing models with Hermes Agent (@NousResearch) - a CodeAct agent that writes Python to call its tools, not JSON function calls. 4 models, RTX 5090, tested under Hermes's real system prompt. ~~ here is the final leaderboard: 🥇 Qwopus-18B — 92.7 🥈 Qwen3.6-27B — 92.4 🥉 Nemotron-Cascade-2-30B — 90.5 4️⃣ Hermes-4.3-36B — 84.3 ~~ no model wins all four axes: - Qwen 27B = perfect multi-step loops + instruction-following, but weakest long-context recall (~70%) - Nemotron + Qwopus = flawless long-context (100%) but worst at multi-step (50%) - Hermes 36B = solid, but OOMs at 64K context on 32GB → that 0 tanks its score the "best agent model" genuinely depends on your workload. ~~ methodology most "function-calling" benchmarks score JSON tool calls. Hermes is code-as-action, which means that the model writes Python. I tested that, under the real ~3.5K-token agent prompt.

@TheAhmadOsman traffic-driven quant downcast. PoS as the inverse of QoS.

Opus 4.8 could be the same nerfed opus 4.6 in 4bit rather than 1.58bit 🤡 I don't trust those clowns Don't waste your money on a Claude Max subscription, they will keep rugpulling you

@LottoLabs @ntbrown01 It is wicked fast, but they need to polish edges around reliability.

@ntbrown01 Interesting

Anthropic is right in the verge of getting lost in the sauce

@LottoLabs 27b-written code will power the software innovation of the next few years. Talk about a model punching way above its weight. Pun intended.

The biggest dark horse in all of ai right now Is qwen 27b on a 3090 going 70+ TPS Literally 6 year old tech that can think

@dakshgup cpu compute that generates training data for subsequent gpu evolution

all coding is just turning gpu compute into cpu compute

@Hikari_07_jp Try @papercliping , helped me regain my sanity. It supports Hermes, which I only use directly for esoteric hand surgeries now.

To maximize throughput, it's necessary to run LLMs in parallel. However, when n increases to around 30, the harness can't handle it. I want to solve this problem on my own, so I'm planning to fork Hermes.

I have 64Gb low cas non-ECC and it’s fine. Color me uneducated, I honestly don’t know why have more RAM (unless it’s Mac). CPU with large cache is much more consequential, X3D edges noticeably on Ryzen builds. Choke is cross-card tensor parallelism over PCIe and not having NVLink. If 6ks had NVLink it would cannibalize a lot of their DC market.

My setup has two RTX RPO 6000 cards. In this case, what RAM do you think would be ideal? I'm planning to upgrade in the near future and I'm torn between 512GB and 256GB. Please share your opinions.

Catching up on latest @theallinpod. It seems like local AI is becoming mainstream. I feel seen.

@YashHustle_22 Codex doesn’t get a TIN nor files 83b.

Can you call yourself a founder if your entire product was built by Codex?

@rohitdotmittal What do you mean by AI? Camera roll, noise cancelling and sms code from messages are AI and we cant function without those.

It’s insane to think that Apple has zero AI execution. Absolutely zero. Nothing they did worked, and they are not even trying. Still, it’s a $4.6 trillion company.

@usr_bin_roygbiv Always run if HM interview is 30 min. This is exactly how much time they’d invest in you.

Getting a job in SF: > recruiter phone call - 15m > hiring manager - 30m > talk to owner/founder - 1h > offer Getting a job in NY: > fill out our workday application > on teams > 3 HR screens > hiring manager shows up hungover > cheat lc > you look like swes ex, don't get job

@ionthedev The best. Answers all the Whys instantly.

The best part of the day as a founding engineer

@kai_xbt Subpoenability of third-party processors without data owner consent is one of the reasons folks deploy local AI. “I’d love to produce logs, your honor, but I physically don’t have them”

Sam Altman reveals there are zero privacy protections for what you tell ChatGPT and says it's "very screwed up" "People talk about the most personal shit in their lives to ChatGPT. Young people especially use it as a therapist, a life coach. Right now if you talk to a therapist or a lawyer or a doctor, there's legal privilege. Doctor patient confidentiality, legal confidentiality. We haven't figured that out yet for ChatGPT. If you talk to it about your most sensitive stuff and there's a lawsuit, we could be required to produce that. I think that's very screwed up. I think we should have the same concept of privacy for your conversations with AI that we do with a therapist. And no one had to think about that even a year ago"