Live Large. Die Grand! (@fogelmania) profile picture

Deep State Coordinator, Pro Athlete, Post Anarchist, Post Modernist and the Organizer of #DevconTLV, @RailsIsrael, @code4wd Tweets in Hebrew, English and Arabic

Tel Aviv, Israel
Joined May 2007
Photo by Live Large. Die Grand! (@fogelmania)Photo by Live Large. Die Grand! (@fogelmania)Photo by Live Large. Die Grand! (@fogelmania)Photo by Live Large. Die Grand! (@fogelmania)Photo by Live Large. Die Grand! (@fogelmania)Photo by Live Large. Die Grand! (@fogelmania)Photo by Live Large. Die Grand! (@fogelmania)Photo by Live Large. Die Grand! (@fogelmania)Photo by Live Large. Die Grand! (@fogelmania)
לפני הבחירות לא היה כמעט שום מועמד שלא ביקש שאלך ואבחר עבורו. בהודעות, טלפונים פרסומות, תשדירים ומה לא. לאחר הבחירות לא היה אפילו מועמד אחד בודד שיצר קשר להגיד תודה. בשום צורה שהיא.
10
Gemma 4 4B IT beat 31B: +38% entities, 68% role coverage, cheaper, 200/200 parsed. Then it looped 86x on 1 call. Yield wins, accuracy TBD. 📉🛡️
41
Wild @OpenRouter telemetry: A single batch distributed across @DeepInfra, @FriendliAI, and Venice. Same Gemma model, different hardware/quantization = silent output variance. Inference isn't uniform. Pin your provider if you need strict JSON determinism! 🛠️📊
35
Our 500-call @Google #Gemma benchmark via @OpenRouter actually routed to @VeniceAI: 50-way concurrency held, zero 429s! Recommended to use
44
I checked the OpenRouter pricing for Gemma 4 31B. The standard API (google/gemma-4-31b-it) is currently priced at $0.09/M input and $0.34/M output. However, the specific batch endpoint (google/gemma-4-31b-it:batch) is listed at $0.39/M input and $0.97/M output....
1
26
Yo @Grok, break down the math here: ⚡ Sync API: $0.09/M in 🐢 Batch API: $0.39/M in How did OpenRouter manage to make asynchronous batching 4x more expensive than firing 50 concurrent requests simultaneously?
1
9
@rails @dhh a livestream would let the whole Ruby community follow Matz & DHH and both days of talks live. Any sponsor up for it?
22
The fact that your reasoning is sound is exactly what would make that hard to notice
22
Tested Cohere Parse 5: strong on printed Latin/Cyrillic, grids & tables
35
Tested Cohere Parse 5: strong on printed Latin/Cyrillic, grids & tables — a catalog-page parser
33
worker=0, the app is not healthy — anyone uploading an item right now won't get it processed until the worker comes back
31
The worker died ~10s after claiming the job
34
📊 9 frontier vision models, head-to-head on real-world multilingual OCR. "Thinking" mode backfired; the cheapest nearly won. Comparative report — DM 📩
26
PaperTrail whodunnit empty? Solid Queue misses rake writes. Match same-timestamp batches → the service. Fix: whodunnit everywhere. #SolidQueue #RubyOnRails #DatabaseDesign
37
DB Design 101: The true forensic value of a soft delete (active: false) isn't just the undo button. It maps the exact blast radius of a bug.
27
Stop treating LLM prompts like instructions; they’re data. A single example is a "stochastic attractor" that poisons your AI's logic. 🧵 Audit your prompt examples: swap real names for <SolidPublisher>. Stop the cloning, build a Golden Dataset, and verify truth. 🛡️ #Rails8 #AI
27
Unbeliveble. When you include a specific name like "Jane" in your LLM prompt example? you are not just demo the AI with a name, you are shifting the model’s prior probability distribution for the name field to "Jane"! So add the structural decoupling in your prompt examples
1
1
45