Peter Banda (@0xbnd) header
Peter Banda (@0xbnd) profile picture

CSO @ https://nitter.cf/t.co/SLWp7LpcdH | builder / thinker 🤖 #EDCWG @AlzheimerEurope, ex @uni_lu, 2 PhDs (AI / ALIFE), 3 kids, dog person

Graz, Austria
Joined February 2023
Photo by Peter Banda (@0xbnd)Photo by Peter Banda (@0xbnd)Photo by Peter Banda (@0xbnd)Photo by Peter Banda (@0xbnd)Photo by Peter Banda (@0xbnd)Photo by Peter Banda (@0xbnd)Photo by Peter Banda (@0xbnd)Photo by Peter Banda (@0xbnd)Photo by Peter Banda (@0xbnd)Photo by Peter Banda (@0xbnd)
Our shift-symmetry paper is out on arXiv 🥳 How many patterns on a wrap-around grid look exactly the same after a shift? We give the exact count - in every dimension, for any grid size and number of colors. With Prof. John Caughman and @tlehman #Combinatorics #CellularAutomata #FormalVerification #Lean4
1
2
1
2
685
Replying to @0xbnd @tlehman
Every numbered statement is machine-checked in Lean 4 (@leanprover): 370 statements, ~30k lines, 3,036 theorems re-checked by the kernel. Then @tlehman audited by hand that the Lean versions of all 27 results say what the paper says - 529 links, 27/27 approved 🙏
1
43
🚀 Minor release 1.7.0 of our #Scala client is out! - Tool loops keep the model's reasoning: @OpenAI, @AnthropicAI, Gemini, Vertex AI, DeepSeek, @MistralAI, MiniMax - Failover policy: per-try timeouts, deadline, parse retries - Case-class schemas: Option = required + nullable
2
3
8
241
Also new: GPT-6.1 Terra, MiniMax-M3.1-Flash, Gemini URL context, Vertex AI global / EU / US endpoints, and the decision models Sarvam Prism + Bespoke Nimble. Nullable tool parameters need a different schema form per host - we measured 7, the client picks the right one 👇
1
108
Sláinte, Dublin! The Irish PM Micheál Martin welcomed us at the #36AEC opening, and the days that followed were moving and inspiring as every year 🧠 Squeezed in a swim in the sea too 🏊‍♂️ Huge kudos to @AlzheimerEurope and the organizers! #EDCWG
1
2
107
And Dublin welcomed us with a Halloween parade the night before - spooky giant included 👏
1
69
🚀 Version 1.6.0 of our #Scala client is out! Meta's Muse, Gemini 4 Argon, DeepSeek V4.1 Flash, Jev 1.14, Liquid's d1-ultra, Perplexity's v1.2 deciders... all live-tested 🔥 @typesafeai @liquidai @perplexity_ai @GeminiApp Release notes in the comments 👇
1
1
2
161
🖼️ 𝐉𝐞𝐯 𝟏.𝟏𝟒 𝐢𝐧 𝐭𝐲𝐩𝐞𝐬 Images, bounding boxes and 262k-token context from @typesafeai's latest - all of it typed: BoxQuestion in, BoundingBox out. Finally, a decision model that can point 😉
1
20
Tested the new long-context decision model on its launch day - 513 questions and 296K tokens in a single call 🔥
Replying to @alex_whedon
Tested ✅ Snap is in our Scala client (ships in 1.6.0): 513 questions in one call in ~1.4 s, where most decision APIs cap at 32-255 🔥 Wishlist: data-URL images, a request-id header, decision-only /v1/models. Max input we hit: 295,957 tokens, not 384K 🤔 github.com/cequence-io/opena…
2
109
🚀 Version 1.5.0 of our #Scala client is out! Decision models are taking over the world - Jev, d1, Luna, Microsoft-Decision-1, Perplexity's decider... so we put eight hosts behind one interface 🔥 @typesafeai @liquidai @OpenAIDevs @MicrosoftAI Release notes in the comments 👇
1
1
1
2
2,398
🧩 𝐓𝐲𝐩𝐞𝐝 𝐝𝐞𝐜𝐢𝐬𝐢𝐨𝐧𝐬 & 𝐠𝐮𝐚𝐫𝐝𝐫𝐚𝐢𝐥𝐬 - decide[T]: case class in, case class out, with probabilities - JSON schema from any case class (Scala 2 & 3), via @JsonSchemaDescription / @JsonSchemaRange - rerank for RAG - guardrails with any decision model
1
1
1
54