One-person team behind SmileToAI. I ship features, read robotics textbooks, and program my own hypertrophy blocks. Warsaw.smiletoai.com Warsaw / PolandJoined March 2018
This week's convergence argument is about text models. I have a small image-model version: 16 first tries across 4 models, 6 misses, and no two misses alike.
A clock asked for 4:35 came back 3:35, about 1:00 and 10:10. Only the 10:10 has an obvious source, the watch-ad pose.
@utsavpm If both products sign in through the same LinkedIn app, check this early. LinkedIn seems to keep one token per app and member, so connecting the second can revoke the first. That took down our publishing once.
@altryne Your 2 and 4 stack against startups. If an agent only gets good after weeks of use, a new one can’t show its value at signup, and people won’t hand weeks of intimate context to a company they don’t trust yet. So the trust has to come first.
@al3rez Those two numbers fit together. With 98.6% of input cached, every turn still re-bills the whole prefix, just at the cache rate, so each extra turn pays again for everything already in context. My guess for delegation: the second model starts with a cold cache.
LiteLLM's PyPI release carried a credential stealer for the second time this year. 1.107.0 (Oct 21) ran on every Python start, even where nothing imports litellm.
A gateway holding every provider key is what attackers want most. One reason we wrote our own adapters for 4 APIs.
@pauliusztin_ So the catalog line ends up being the whole retrieval step. If a skill’s description doesn’t sound like the task in front of the agent, it usually won’t open the SKILL.md, and a lot of the tuning moves into those short descriptions.
We added a test that fails if our Alembic migrations have two heads. In the four days after it landed, main split into two heads four times.
Each PR had one head on its own branch, so the test passes each one. The split only exists after two merge, and nothing checks that.
@0xPascual Active params set the compute per token, not the memory. All 126B of the audio-flash weights still have to sit somewhere, about 63 GB even at 4-bit. tts-tiny (7.9B) and the 8.6B ASR fit a 24 GB card, flash doesn’t.
@JoePro Counting’s a good one, though seven dice can’t all show different faces. Six can, and every face readable is a test on its own. Fair point on tries too, two each would show who just got lucky.
Same prompt, 4 models #4: a mirror reflecting a clock at 3:00. Only Nano Banana 2.1 and gpt-image-2 flipped the numerals and put the hand at 9. FLUX.2 Pro flipped some numerals, not others. Grok Imagine drew a normal clock. One try each. What should #5 test?
Asked a model to turn a 21,346-character doc into a read-aloud script, every word kept. It gave back 2,765 characters.
So no single call re-emits ~5k tokens now: 2,000-char sections, joined in code, no compose pass. A 21,276-char test doc keeps 19,855 of 20,816 spoken chars.
@mardehaym The item about not trusting the agent’s own status report goes one layer down. Our worker took a provider’s “succeeded” at face value even when outputs came back empty, and would have billed for nothing. Now a video or image run only settles once the file is stored.
@riya_mishra007 I'd start by finding which requests make up the bill. In chat it's usually long conversations, since every turn resends the whole history. Caching discounts the repeat; summarising older turns removes it, and keeping the last few verbatim is what protects the UX
@buildwithsid That $280k prices all 70B as fresh input. On an agent loop most of that is usually cache reads, which Opus 5.5 bills at $0.20 per million, a twentieth of the input rate. At 90% cache the input side is nearer $40k.
6 Followers 114 FollowingConnect your agents, delegate anything, track everything. Any-vendor AI agents work as a team, and every handoff gets a receipt. Built by @midwit.
297 Followers 469 FollowingEntrepreneur & Indie dev👨💻, AI & Computer Vision freelancer, currently building https://t.co/LZHbrxIgHz 📲 Follow for original content about tech, product & life
78 Followers 1K Following"ingeniero" de sistemas, filósofo entusiasta. apasionado por la buena comida y la simplicidad o el caos de resolver un buen problema. | 🇨🇴
181 Followers 2K FollowingYour Den of Darkest Desires.
Effortlessly craft your wildest fantasies into stunning visuals.
Join now @ https://t.co/k3khBYLpgg
157K Followers 79 FollowingSane + 🌶️ takes in an insane AI world... AI capabilities researcher: co-created RLHF/ChatGPT @ @openai now trying to right the wrong 🤭 (ceo @typesafeai)
1.8M Followers 2 FollowingWe're an AI safety and research company that builds reliable, interpretable, and steerable AI systems. Talk to our AI assistant @claudeai on https://t.co/FhDI3KQh0n.
4K Followers 192 FollowingTürkiye Cumhuriyeti Varşova Büyükelçiliği Resmi Hesabı / Official Account of the Turkish Embassy in Warsaw / Oficjalne Konto Ambasady Turcji w Warszawie
1.9M Followers 1K FollowingCo-Founder of Coursera; Stanford CS adjunct faculty. Former head of Baidu AI Group/Google Brain. #ai #machinelearning, #deeplearning #MOOCs
5.5M Followers 4 FollowingOpenAI’s mission is to ensure that artificial general intelligence benefits all of humanity. We’re hiring: https://t.co/dJGr6LgzPA