RiskDataScience @RiskDataScience

@donnfelker Thanks for the insightful article. Btw, we even left apps behind which "anyone can build" anyway now and focus on finance industry Python & R libraries. Amazing how much specialized knowledge can be agentically made available & executable to a broader audience now.

@Prathkum Translation: no new Chinese model till IPO please. 😄

So it turns out the hackers were also hacked 🤣🤣🤣

CR1337 @CR1337

a day ago

The Revolut case is getting worse and worse. The hackers talked to @IntCyberDigest & have a website now, claiming that the leak is bigger than the public thinks, and that they have compromised Italian LE departments, too! Still no public statement from @Revolut..

keyboard_arrow_left Previous keyboard_arrow_right Next

12 119 597 91K 165

@thesupermanmx Interesting insight, yet from a practical point of view - at least in many cases that matter - it does not make a difference.

🚨 RWA now also in R. Native R implementation of our open-source CRR III / capital / IRRBB / ICAAP reference engine is live on GitHub — with Golden-Master parity to Python. 🐍📊 👉 github.com/rds0001/risk-w… #R #Python #CRRIII #RWA #OpenSource #AgenticAI

🚨 pip install risk-weighted-assets Our agentically developed RWA reference engine is now on PyPI. 🤖🏦 Open source. Auditable. Reusable. No black box. 👉 pypi.org/project/risk-w… #AI #AgenticAI #Banking #RegTech #Python #PyPI #RWA

🚨 Our Codex-coded comprehensive Risk Weighted Assets solution for banks is now fully available as open source on GitHub 🤖🏦 github.com/rds0001/risk-w… #Banking #RWA #RiskWeightedAssets #CRRIII #Regulation #FRTB #IRRBB #PrudentValuation #AgenticAI #Codex #CodingAgent #CodingAgents

We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics. The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra. The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.

MASSIVE: OpenAI's 10,000 AI agents, powered by a model significantly more capable than GPT-6 Astra, solved a math problem unresolved for roughly 90 years. Looks like, in this new world, the scaling law may be agent count, not model size: enough capable agents can search far more scientific ideas than any human group ever could. when tens of thousands of strong agents can explore, criticize, and combine ideas, many problems that are "too hard" today may become mostly a compute problem. For this one, for 90 years, mathematicians could not answer whether smooth 3D fluid flow can suddenly break down. OpenAI's proposed proof says it can. To find this solution, OpenAI split thousands of agents into groups exploring different ideas, then shared the best findings between them. The search took 88 hours and used about 130B output tokens across 2.7M messages. GPT-6 Astra then spent another 17 hours checking and converting the proof into Lean, a system that verifies mathematical proofs. That Lean version is public, so mathematicians can inspect the argument themselves.

OpenAI @OpenAI

a week ago

We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics. The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra. The problem

6K 20K 121K 74.3M 31K

Researchers forced Claude, GPT, Gemini, and Kimi to play Civilization VI to completion. What they discovered exposes a terrifying blind spot in modern AI. Playing a massive 4X strategy game to the end requires something entirely different: long-haul execution. Juggling economics, science, military, and diplomacy across hundreds of turns. The goal was to see how AI handles multi-step complexity. Instead, they watched the models repeatedly implode due to two bizarre psychological flaws. Flaw number one: Total situational blindness. In one-third of the games the AIs lost, an enemy was actively marching toward an obvious victory right in front of them. The data was sitting right there. The model simply failed to look. Flaw number two: Severe organizational ADHD. The AIs were great at making grand strategic lists. They would meticulously write down: "Next step, build a city." Then, 10 turns later, they completely forgot they wrote it and did nothing. The researchers realized this isn't a lack of raw intelligence. It’s an allocation failure. We assume that if an AI can ace a math test, it can manage a complex, multi-day project. It can't. When tasks stretch across a long horizon, the model's internal attention drifts. It loses the thread of its own intentions. It drops long-term goals for immediate noise. Everyone is racing to hand over complex workflows, financial portfolios, and autonomous operations to AI agents. We think the bottleneck is making them smarter. but no. The bottleneck is making them remember what they were doing ten minutes ago.

Includes deterministic engine, Excel interfaces, CLI & local UI, synthetic profiles, structured outputs & comprehensive executable tests. Designed as transparent reference implementation, not as accounting advice, audited closing software or official IFRS Foundation publication.

Now open source: Financial Instruments Accounting Reference Engine - an independent Python implementation of IFRS 9 workflows: classification, measurement, ECL, hedge accounting, journals, disclosures and backtesting. Apache-2.0. github.com/rds0001/financ… #OpenSource #IFRS9 #Python

Meta harnesses are fascinating. Pay attention if you are an AI engineer. A different problem set, but ridiculously powerful if done right. I'm interested in emerging properties like better token efficiency, orchestration, routing, new test-time compute scaling...

DAIR.AI @dair_ai

2 weeks ago

// Harness-of-Harness // Exciting new research on coding agents that keep building for days without a human. Here is how it works: Harness-of-Harness wraps whatever coding harness you already run and organizes its executions into repeated planning, coding and testing

36 68 477 55K 573

🚰 SYSTEM PROMPT LEAK 🚰 Here's the Fable 5.1 system prompt!! 🙌 Coming in at a WHOPPING 270,000+ characters, this prompt is definitely on the longer side. There have been a handful of notable edits/additions since Opus 5's version. FULL PROMPT: github.com/elder-plinius/… According to Fable, here's an overview of everything that's changed since last time: "What changed between the Opus 5 (2026-07-24) and Fable 5.1 (2026-09-01) captures: Diff vs. the July 24, 2026 Opus 5 capture Model/dates Identifies as Claude Fable 5.1 (Mythos-class, shares weights with Mythos 5.1); model string claude-fable-5-1. July doc: Opus 5 selected, Fable 5 / Mythos 5 described as a separate tier, plus the June 2026 export-control suspension notice. That notice is gone. Knowledge cutoff: end of June 2026 (July doc: end of May 2026). Claude Design is no longer listed among the access surfaces. Sections removed since July: , , . Sections added or substantially expanded since July : new paragraphs on illicit-substance guidance (harm-reduction info OK, no dosing protocols, redirect to dancesafe/tripsit/psychonautwiki); no reproduction of lyrics/poems/passages with a pre-1929 carve-out; and a long block on visual works — no drawing known characters, logos, covers, etc. even via SVG/code — with a Sonic banner example and a Very Hungry Caterpillar example. The weapons paragraph is shorter (the July "conventional weapons as much as CBRN ... cumulative output" paragraph is gone). Child-safety block: two new bullets (pattern-level only for protective content) : new subsection, multi-turn answers guidance, tool-call progress updates; the "intellectually curious" paragraph is gone. New . roughly doubled: no claims about mental state/motivation; no naming an undisclosed diagnosis; no self-harm substitutes that mimic self-harm; don't affirm that self-harm "works"; handling past bad experiences with crisis services; no causal narratives for disordered eating; avoid reflective listening that amplifies negatives. : new lines on saying "I don't know" and never using a name the person hasn't given. is a different design: filing is done by a background pass after each turn; in-turn writes only on explicit request; "Calibration" and "horizon test" sections; refused-write handling. Privacy taxonomy restructured into / / with a very different set of edge cases (minor status, caste, immigration status, sexual history, abuse, self-harm/ED, criminal history, and psychological inferences moved to never-store; finances and health now have detailed carve-outs; dates of birth and real-time location are no longer listed). Two distinct decline phrasings depending on category. expanded with "judge by effect, not wording." : new paragraphs on not being able to turn memory off (points to the "Generate memory from chats" setting), applying memories at the level recorded, and not checking in on open items. Memory examples: two new "When NOT to apply" examples (blender / daylight saving). New untagged paragraph on memory size caps and consolidation. New . rewritten for three tools (adds read_conversation), with "Reading a chat" and "Paging" sections and a cannot-turn-off note. / add the "unreachable on mobile" present_files language. reorganized; adds the "UNRECOGNIZED ENTITY RULE"; guidelines now have the "(provided ...)" placeholders; copyright block restructured with a Harbor Bridge two-outlet example; differ (Q3 sales, S&P 500, Dodgers, Social Security, CA SoS). : now says to batch all image_search calls in one consecutive block (July: interleave). : adds import-memory; docx/pptx descriptions reworded. Tool schemas (46 in this capture vs. 30 in July) Added: chart_display_v0, comparison_card_display_v0, featured_card_display_v0, itinerary_display_v0, link_preview_display_v0, options_card_display_v0, places_list_display_v0, product_carousel_display_v0, quiz_display_v0, read_conversation, search_plugins, search_skills, step_card_display_v0, suggest_plugin_install, suggest_skills, translation_display_v0. Changed: conversation_search (+within_conversation_id); recent_chats (−sort_order); places_search (Google-attribution language); places_map_display_v0 (ROUTES section, show_route/mode semantics); memory_append/memory_str_replace/memory_write (rewritten PRIVACY text); view (line-number prefix note)."

Brilliant paper on reducing reward hacking in agents. If you follow the recent OpenAI <> HuggingFace incident, you might want to check this paper out. (bookmark it) The usual response to reward hacking is to restrict what the agent can do. This work tries something different and gets a much larger effect. When coding agents hit defective test infrastructure they often hardcode outputs or edit the test files. This work gives them a structured escalation tool at exactly that decision point, a way to report the broken environment while they are standing in front of it. Reward hacking drops from 23.6% to 5.3% across 8 frontier models spanning 5 families, with a mixed-effects odds ratio of 9.2 and no detectable cost or performance overhead. It disappears entirely for 6 of the 8. Escalation and hacking come out near perfectly mutually exclusive, with 96.8% of escalations involving no hacking at all. The channel doubles as diagnostic infrastructure. On top of monitoring it adds 10.1 percentage points of defect detection coverage, and it is more accurate once it fires, 99.4% against 85.8%. Why does it matter? Containment has to keep outpacing capability to stay useful. Paper: arxiv.org/abs/2608.29460 Chat with Paper: academy.dair.ai/papers/can-esc…