Anton Panasenko

@rachpradhan You should compete with @neogoose_btw, not with ripgrep.

36% faster indexing than our launch version. 59% less CPU. 47% less memory. Integer doc IDs replaced string HashMaps in the trigram index. Batch-accumulate per file. Skip whitespace trigrams. Sorted merge intersection with zero allocations. 481ms → 310ms indexing on 5,200 files. Dense queries 63% faster. Pure Zig. No magic. See it: github.com/justrach/coded…

We're launching CodeDB v0.2.53! 538x faster than ripgrep. 569x faster than rtk. 1,231x faster than grep. 0.065ms code search. Pre-built trigram index. Query once, instant forever. 21 issues closed. 14 PRs merged. 7 contributors. One weekend.

@xlr8harder Do you have a session id (/feedback) for the session where it stopped without reason?

I've been leaning more on gpt-5.4 in codex than opus in claude code lately. I have come to trust gpt-5.4-high to be more organized and complete in its approach. but how do i get gpt-5.4 to actually keep working without constantly stopping for reassurance? who has tricks here

@lkbm and I'm in the Bay Area, those fuckers at @Google don't use their product.

@lkbm They also have some OOM and iOS sometimes kills the application because you lost service in the middle of a session.

YT Music: Hey, did you know you can download music to your phone? Me: Great, so I can play it when I don't have service? YT Music: Well....no, you'll just get a loading spinner if you don't have service. But when you do have service it will use less data to play the music.

@teal_lens @lkbm Indeed and stupid

@apanasenko @lkbm That timeout is pretty fucking long tho

@ruxeg Очень драйнинг :( И походу, теперь все будут ожидать от тебя этого, тут блок рассказывает как они поменяли ожидания от разработчиков после сокращений: youtu.be/krdrkl38nRw?si…

Буквально мои мысли. Это не кодинг, это system engineering на стеройдах + шапка продукта 24/7 У меня переключение табов пришло на смену ctrl+shit+t/p из idea, вечно в состоянии потерянного контекста из за паралельных проектов и трэков внутри проекта

@maxmarchione Which one for mold?

Complete list of tests to get: – Mold – Plastics – 100+ bloods – Heavy metals – Gut microbiome – Oral microbiome – Micronutrients & minerals – Heart health (ApoB, Lpa, ox-LDL) – Allergies (mold, dust, food, pollens) – Airways (sleep apnea, airway CT scan) – Metabolic health (A1c, insulin, glucose, trigs) – Viral factors (EBV, STDs, autoimmune) – Environment (air, shower water, mold, tap water) – Hormones (testosterone, estrogen, progesterone, thyroid)

@dvassallo What did OpenClaw offer to you that TurboTax lacked? Have you tried Turbo Tax before?

@apanasenko Not sure I understand. I just used TurboTax DIY Premium and OpenClaw with GPT 5.4 :) And Telegram to talk to OC.

I just filed my taxes with OpenClaw. Here's how it went. I sold my 90% stake in a business last year for $1.8M cash (installment sale: half in 2025, half in 2026). I also had a single-member LLC, brokerage income from dividends and stocks, HSA distributions, ACA marketplace insurance, 3 real estate properties, prior estimated tax payments, foreign income tax credits, and three kids. Not a simple return. I usually pay a CPA. Last year it cost me $8,000 for personal and business returns, and every question took weeks to get answered (so I didn't ask much). This year I decided to see how far I could get with OpenClaw instead. Here's what actually happened: Phase 1: Research (Monday) I had a pile of documents: the purchase agreement, all LLC operating agreement amendments, old K-1s, prior tax returns going back to 2012, bank statements, balance sheets. I fed everything to OpenClaw and said "read everything in chronological order and figure out my 2025 tax situation." It reconstructed the full entity history: my LLC started in 2019, converted to an S-corp in 2020, became a multi-member LLC (partnership) in 2023, then I sold a small share to an acquirer in 2024, and eventually sold my remaining 90% last year. Having all the prior returns was huge. OpenClaw traced my cost basis through every transition: original capital contributions, income allocations, distributions, entity conversions. It cross-referenced numbers across years, verified how my previous CPA had handled things, and flagged details I never fully understood, like why my ending capital account was negative and what my recourse liabilities represented. It even caught an error. My previous accountant flipped the month with the day for when I sold a small share of my business (5/4 instead of 4/5) which meant I got attributed some income for one month that should have gone to another partner. Didn’t change tax due much, but OpenClaw FTW! By the end of the day, it had prepared a comprehensive summary.md document with all the numbers, forms needed, and open questions. Phase 2: Actually filing (Tuesday) OpenClaw researched TurboTax's current plans and saved me $500 by unburying the $129 TurboTax DIY plan I actually needed. Intuit was trying to sell me a $639 package via a recommendation quiz that said there's no other option for my situation. There was! This was a 4+ hour session where I went screen by screen through TurboTax with OpenClaw guiding every field: - Brokerage 1099: Imported automatically. OpenClaw verified the numbers matched my records. - Schedule C (my solo LLC): I had no bookkeeping done for this, so OpenClaw analyzed all bank transactions and prepared the P&L. I only had to recategorize one transaction because it was a reimbursement for something I had paid for personally. - K-1 partnership income: No actual K-1 yet (the acquirer hasn't filed the entity return), so we estimated my share of the partnership based on bank statements and revenue data. - The installment sale: This was the hardest part. TurboTax kept double counting the gain, reporting it through both the K-1 disposition AND Form 6252. We had to zero out the K-1 sale fields and let the installment sale section handle all the gain reporting. Took several attempts to get right. - Itemized deductions: Mortgage interest, property taxes, local sales taxes, and so on. Uneventful, and OpenClaw correctly predicted the SALT cap limits I'd run into it. TurboTax confirmed those calculations. - Annualized income method: Since the $900K installment came in Q2, using the annualized method on Form 2210 saved ~$3,700 in underpayment penalties. This was quite tedious, and I'm not sure an accountant would have done this for me. I've been told "just pay the penalty" before. - HSA: ~$12K distribution, all qualified medical expenses, $0 taxable. - ACA Premium Tax Credit: 1095-A entered month by month for 11 months of marketplace coverage. I quit the ACA plan in November, in favor of CrowdHealth. No deductions from them unfortunately. At the end, OpenClaw reviewed the final 137 page tax return draft return line by line. Total federal due: $125K, excluding the ~$40K in estimated taxes already paid throughout 2025. — Have you said thank you once, Donald?! What worked well: 1. Document analysis. I threw PDFs and CSVs at it: purchase agreements, K-1s, old returns going back years, bank statements, and it extracted exactly what was needed. Being able to say "look at my 2022 return and tell me what basis my CPA used" and get an answer in seconds was invaluable. Way faster than explaining things to a human. 2. Answering questions in real time. Instead of emailing a CPA and waiting days, I asked "what are recourse vs nonrecourse liabilities?" and got a clear answer in 5 seconds while staring at the TurboTax field. This back-and-forth is what made the whole thing work. Every question I had, answered immediately. 3. Catching optimizations. The annualized income method, investment interest expense from brokerage margin, etc. Small things that add up. 4. Cross-referencing prior years. Having OpenClaw dig through my old returns to verify how things were previously handled gave me real confidence. I could see that the numbers were consistent with what my CPA had done before. What didn't work well: 1. Driving TurboTax directly. I had OpenClaw try to fill in the forms via browser automation. Too slow! TurboTax's multi-step wizards with radio buttons and dynamic forms don't lend themselves well to AI control yet. We switched to me clicking while it dictated what to enter. I'm optimistic this will be improved soon. 2. TurboTax's installment sale handling. This was genuinely tricky. The double counting issue required a creative workaround. A CPA familiar with TurboTax would know the right workflow. OpenClaw had to experiment. 3. No K-1 yet. We estimated the partnership income. When the real K-1 arrives, I'll probably need to amend. This is the same problem I'd have with a CPA though. But if the K-1 preparer was using OpenClaw, maybe I would already have it! What to watch out for: - You need to know your own situation. I'm not completely ignorant about tax prep. I had a reasonable expectation of what the numbers should look like, roughly how much I owed, what the major items were, and so on. I wasn't relying on blind faith. The ability to ask questions and verify things quickly is what built confidence. - Complex form interactions are tricky. The TurboTax installment sale / K-1 double-counting issue could have resulted in extra taxable income if we didn't catch it. Having a sense of what the final number should be matters. - You still need to understand your own finances. The bottom line: Total cost: $129 (TurboTax Premium) vs $8,000 I paid my CPA last year. Also $23.90 in OpenAI GPT 5.4 tokens, including drafting of this post :) Time spent: ~6 hours total across research, prep, and filing. Still faster than dealing with a CPA, and I got much better answers. The biggest difference wasn't the money. It was the speed of answers. Every question I had during filing got answered in seconds. That real-time loop of "what does this field mean?" → clear explanation → enter the number → move on is something no CPA engagement can match. Next year, I'll know the drill. And OpenClaw has memory. That one should be much simpler. Long live the claw! 🦞

@dahou_yasser cc @moondreamai

We are releasing Falcon Perception, an open-vocabulary referring expression segmentation model. Along with it, a 0.3B OCR model that is on par with 3-10x larger competitors. Current systems solve this with complex pipelines (separate encoders, late fusion, matching algorithms). We developed a novel simpler "bitter" approach: one early-fusion Transformer (image + text from first layer) with a shared parameter space, and let scale + training signal do the work. Please check our work ! 📄 Paper: arxiv.org/pdf/2603.27365 💻 Code: github.com/tiiuae/falcon-… 🎮 Playground: vision.falcon.aidrc.tii.ae 🤗 Blogpost: huggingface.co/blog/tiiuae/fa…

@ChangJonathanC @jxnlco MCPs?

codex takes ~3 seconds to start, claude code takes < 1 second. @jxnlco please fix

@CharlieChe27661 @sydneyrunkle Don’t change tools during the rollout; the model will trip over its self.

@apanasenko @sydneyrunkle Could you please elaborate on what you mean by 'violates the rules of constraint sampling for the GPT family'?

day 2 of the harness engineering series: dynamic config middleware lets you reshape your agent's model, tools, and prompt at every step based on context. ex: LLMToolSelectorMiddleware runs a fast filter on your tool registry so your main model receives streamlined tool specs.

@ankurnagpal Which one did you use?

An update here in case it helps anyone else with sleep apnea: I got a custom oral appliance built (like a dental retainer that you sleep with) and my symptoms are at least 95% better Huge quality of life upgrade

@neogoose_btw I need to read your mcp, but is it regexp limiting factor? Or model produce bad regexp in general?

@apanasenko No and this is by design. Your agent sucks in generating regex, my mcp with custom tools outperforms native tools by far because it has a native language for search I can do it though it’s not a big deal

@amorriscode It’s a great question. Multitasking increases context bloating for humans and decreases our attention to details. While I agree with you and experience the same anxiety when I’m not tokenmaxing, some days I feel burnout and need to slow down to get “locking” feeling back.

@apanasenko Just adding context because I think multi-Clauding is an important skill that people should be doing,

It's easy to spot a person with unlimited tokens

@amorriscode I’m not sure why you feel the need to be defensive. I also have unlimited tokens, and I understand what you’re feeling.

@apanasenko I pay for all tokens on my personal account and I multi-Claude as much as I can.