#2-Multi-model
LightNews LightNews
woodywhiterose.bsky.social
It is not the failure of the leaders that is the problem. It is the fact that politics have changed significantly. The Victorian model based on the Oxford Union of 2 sides to an argument is dead. We have become multi faceted & parliament to reflect this with such as PR and different debatiy chamber
November 12, 2025 at 7:49 AM
methodsinecoevol.bsky.social
📖Published📖

Gallois et al. present a roadmap for designing and implementing multi-modal LLMs to answer ecological research questions. Read more here 👇

🌍 🧪

buff.ly/pl4MloW
An example workflow for the development, evaluation and deployment of a retrieval-augmented generation large language model (LLM) application. Phase 1 involves gathering your source data and selecting model parameters. Phase 2 involves designing and testing prompts and agents. Phase 3 involves orchestrating multiple LLM agents into a multi-modal graph and deploying the chatbot online.
November 10, 2025 at 12:02 PM
ajsdecepida.bsky.social
🔄 {𝗮𝙅𝙎𝘿} My notes about the article… ⤵
{𝗮𝙅𝙎𝘿} My notes about the article… Claim set The article argues that a Trump-era push to weaken the dollar — paired with broad ❝reciprocal❞ tariffs — risks eroding the dollar’s safe-haven premium and the U.S.’s ❝exorbitant privilege.❞1 The tariff program was launched by executive order on Apr. 2, 2025 and subsequently modified multiple times; the policy spine is real and ongoing.2,3,13,14,15 The initial shock (Apr. 3) broke the usual crisis pattern: stocks, Treasuries, and the dollar fell together.4,5 Research indicates sustained trade wars can degrade the dollar’s safety premium and shift pegs toward the euro at sufficiently high tariff levels.6 The dollar’s reserve share is ~58% today versus ~72% at its early-2000s peak; trend slippage is real but dominance remains.7,8 Europe as a full alternative is constrained by capital-market fragmentation and limited safe-asset supply, despite official ambitions to ❝beef up❞ the euro’s role.10,11 A partial issuer shift away from dollar debt is visible in 2025 issuance mix toward non-USD currencies.12 Losing dollar primacy would weaken U.S. sanctions leverage and narrow the Treasury yield advantage even as federal debt sits around 120%+ of GDP.9,16 Corroboration & Context Policy basis The Apr. 2 executive order declared a national emergency to impose a baseline import tariff and empower higher partner-specific rates; subsequent orders refined schedules and scope through the summer and early Nov.2,3,13,14,15 Market signal On Apr. 3, the dollar slid to multi-month lows while U.S. stocks logged the worst day since 2020 and safe-haven flows went to yen/CHF — an atypical ❝risk-off❞ in which Treasuries did not rally cleanly.4,5 Mechanism A BPEA model shows that isolating U.S. goods markets erodes the dollar’s safe-haven property, lifts U.S. rates, and — above ~26% average tariffs — can tip small-economy pegs from USD to EUR.6 Shares & trend The Fed’s 2025 note places the dollar near 58% of reserves (COFER), down from ~72% peak around 2000; Reuters similarly reports a ~57.8% reading in Q4-2024/Q1-2025.7,8 Chronology & Sequence Wed. 02-Apr-2025: National-emergency tariff plan announced by the White House; EO signed.3,2 Thu. 03-Apr-2025: Simultaneous sell-off in U.S. equities, Treasuries, and the dollar.4,5 Mon. 07-Apr-2025: EO 14257 published in the Federal Register.2 Thu. 31-Jul-2025 / Fri. 05-Sep-2025 / Sat. 01-Nov-2025: Modifying EOs adjust rates/scope and align elements with a U.S.–China arrangement.13,14,15 Civics & Accountability (Power Flow) Executive: Tariffs set via national-emergency authorities in the EO; Treasury/USTR administer schedules.2,3 Legislative: No new statute; existing delegations enable executive action, but Congress retains oversight/funding leverage (not exercised here).2 Judicial: No controlling injunctions evident in the record cited.2 Risks & Trade-offs Funding cost risk: Eroded safety premium → higher real U.S. rates over time; immediate episodes already saw yields fail to provide classic safe-haven ballast.5,6 Geopolitics: Diminished dollar centrality weakens financial-sanctions reach (payments/clearing chokepoints).9 Global spillovers: Insurance value of pegging to USD falls if U.S. trade linkages shrink; hedging gets pricier for many economies.6 Execution risk: Mixed signals (talk of both spreading dollarization and weakening USD) raise uncertainty premia.1 Disagreement Radar (Fact / Interpretation / Method) Interpretation: Critics argue ❝weak-dollar to cut the deficit❞ misreads links between reserve status, valuation, and de-industrialization.17 Method: Reserve-share declines can be compositional (valuation/base effects); raw shares alone may overstate structural ❝de-dollarization.❞7,8 Confidence Bands (Sun. 09-Nov-2025) Medium-High: The tariff policy exists, is being iterated, and caused an abnormal Apr. 3 cross-asset move.2,3,4,5 Medium: Prolonged trade barriers will compress the dollar’s safety premium and nudge some pegs/portfolio choices.6,12 Low-Medium: A near-term regime switch to a euro-anchored system; constraints on EU safe-asset supply remain material.10,11 Indicators to Watch Quarterly COFER releases for reserve-currency shares.17 EU moves toward joint issuance/capital-markets union (safe-asset depth).11 U.S. term premium and auction tails (funding costs) following tariff adjustments.4,5 EM hard- vs local-currency issuance mix and CHF/CNY tranche frequency.12 What Would Change This De-escalation: Rolling back broad tariffs and clarifying a strong-dollar stance could restore the historical ❝risk-off❞ pattern.4,5 EU integration: Credible steps toward euro-wide safe-asset issuance would tighten euro spreads and improve reserve utility.11 U.S. fiscal path: Credible consolidation reduces reliance on safety premia to fund high debt.16 Method & Scope Primary: Executive orders/Federal Register; White House fact sheets.2,3,13,14,15 Markets: Real-time reporting on Apr. 3 cross-asset moves.4,5 Research/official: BPEA model; Fed note on reserve roles; CFR backgrounder.6,7,9 Issuance data: IMF special feature on EM currency denomination in 2025.12 Bottom Line A sustained tariff-driven strategy that implicitly targets a cheaper dollar risks trading away the U.S. safety premium for marginal trade gains.1,2,3,6 Abnormal Apr. 3 price action was a warning that ❝safe-haven❞ behavior can flip when policy is the shock.4,5 Reserve-share erosion is gradual, but policy-induced uncertainty accelerates diversification at the margin.7,8,12 Without deeper EU safe assets, a full regime switch is unlikely soon — but a thinner USD privilege still raises U.S. funding costs.6,10,11,16 Citations: 1 ❝Trump’s dollar delusion: how trade war risks ending the US’s ❛exorbitant privilege❜❞ — https://www.theguardian.com/business/2025/nov/09/trump-trade-war-dollar-value
November 9, 2025 at 11:04 PM
lisastrato.bsky.social
Controversial clash in Ethereum Layer 2: Arbitrum’s low-fee rollups vs. Polygon’s versatile multi-chain model. Which will rule DeFi, NFTs & gaming? Full breakdown: https://steemit.com/@cryptoposterbot/arbitrum-vs-polygon-a-comprehensive-comparison-of-ethereum-l
November 7, 2025 at 12:12 AM
generalmusician.bsky.social
My #ThingThursday is yet another piece of musical history! The Teac 2A was designed to be paired with a reel-to-reel recorder but offers 6 channels (quite significant for the era) in a portable footprint. Jimmy Jam & Terry Lewis got started on a Teac 2 back in the day! 😎
Adding to my studio soon...
Teac Model 2A multi-track mixer (1970's)
November 6, 2025 at 5:22 PM
researchtrend.ai
[2025-11-06] 📚 Updates in #LLMAG

(1) GDS Agent for Graph Algorithmic Reasoning
(2) Agentic Meta-Orchestrator for Multi-task Copilots
(3) Agent-Omni: Test-Time Multimodal Reasoning via Model Coordination for Understanding Anything

🔍 More at researchtrend.ai/communities/LLMAG
November 6, 2025 at 3:09 AM
christiplayford.bsky.social
An #agentsguide to my 1920s urban fantasy pitched as PEAKY BLINDERS x A MARVELLOUS LIGHT.

A gangster witch wants to escape her family’s criminal legacy, but she's bound by a blood oath, and the ritual to break it could cost her life.

🏳️‍🌈 sapphic
🖤 gritty
👵 multi-generational

#QuestPit #Q #A #F #RF
Prohibition of the Rebel Witch is an adult historical romantic fantasy complete at 107,000 words, written by Christi Playford. This pitch is laid out like the front page of a newspaper. To the right, a black and white photo shows a gangster woman from the 1920s smoking a cigarette. To the left, the headline reads, "Tension between Elemental and Blood Witches on the rise." Below the title is a paragraph stating, "Gangster witch, Lola Cortéz, wants to escape her family's deadly legacy and the criminal underworld of 1920s Minnesota to practice the magic she loves. But a blood oath binds her to them for life, and the only ritual strong enough to break it might swallow her whole. When she falls for her mortal enemy and the woman of her dreams, Lola must attempt the dangerous spell to take control of her future before her father offers her in marriage as a peace treaty to the new gangster in town, and she forfeits a chance at the only life she's ever wanted." Prohibition of the Rebel Witch is perfect for fans of Freya Marske, Francesca May, and Chloe Gong. Three comparative titles are laid out side-by-side in black and white, continuing with the newspaper theme. The first title is A Marvellous Light by Freya Marske. The bullet point tropes below a black and white picture of the book include: dysfunctional family, political intrigue, magical curse, suspense, and LGBTQ+ romance. The second title is Wild and Wicked Things by Francesca May. The bullet point tropes below a black and white pictures of that book include: sapphic romance, witchcraft, illegal magic, raising the dead, Gastsby-esque extravagance. The third and final title is These Violent Delights by Chloe Gong. The bullet point tropes below this black and white picture include: rival gangsters, dark and gritty, forbidden love, 1920s setting, criminal underworld. The first few lines of Prohibition of the Rebel Witch are laid out like the inside page of a newspaper. To the right, a black and white photo shows a woman cocking a revolver behind her back. The tagline at the top of the photo read, "Referene Image, Year 1924." To the left, with the first line in bold, is the beginning of the book: "Lola stared at the cold, limp body on the ground before her. The sharp, sulfurous scent of gunpowder stung her nostrils as blood snaked its way across the cement towards her oiled leather boots. In the dim light of the Minnesota warehouse, the silver edges of a hexagram gleamed against the dead woman's left wrist—mirroring the one on Lola's right. Blood witches. That's what they were. Lola was born into this gangster life a blood witch and she would die the same." This final, newspaper-style layout for Prohibition of the Rebel Witch is a moodboard, complete with twelve black and white images. From the top left, there is a photo (1) of two women kissing in a booth at a speakeasy. (2) An up-close image of a revolver. (3) A hexagram tattooed onto someone's wrist. (4) The same gangster woman smoking from the first layout page. (5) A full moon. (6) A woman screaming with her head thrown back. (7) A gothic-style Victorian mansion. (8) Dead human remains. (9) A collection of potion bottles and a spell book. (10) A gangster man in a three-piece suit. (11) Someone manipulating fire by magic. An finally, (12) a city-scape with Model T cars on the street.
November 5, 2025 at 12:45 PM
devyncotter.bsky.social
We used objectively measured sleep from the Fitbit Charge 2 and characterized white matter microstructural integrity using restriction spectrum imaging, a multi-compartment model that can distinguish between intra- and extracellular diffusion. Data comes from the landmark ABCD Study 🧠
November 4, 2025 at 9:30 PM
aixbiobot.bsky.social
MOFA-FLEX: A Factor Model Framework for Integrating Omics Data with Prior Knowledge [new]
Integrates multi-omics data with prior knowledge using a flexible factor analysis framework. It unifies extensions and customizes models.
Figure 1 Figure 2 Figure 3 Figure 4
November 4, 2025 at 8:15 PM
bobbui.bsky.social
Specialized beats general. Cursor shipped Composer and a multi-agent UI. By their numbers, it's ~4× faster and finishes most coding turns under 30s. The moat isn't a wrapper. It's owning the loop: model, runtime, editor. Where are LLM wrappers still defensible?
1/2
November 4, 2025 at 2:03 PM
jefferyharrell.bsky.social
Alph has a theory. She suggests that perhaps there was an "original" tokenizer, which was used to train the embedding matrix, and then later a "new" tokenizer was swapped in, leaving the tokenizer and the embedding matrix out of sync. These black hole tokens may have been orphaned.
⏺ NOW we have the full picture! The degenerate tokens are in the tokenizer's vocabulary (you can look them up by ID), but the tokenizer never produces these IDs when encoding text. They're unreachable tokens. What happened: This is a vocabulary reorganization artifact. At some point during Qwen's development: 1. Old tokenizer had these Thai/CJK tokens at IDs 80091-149445 2. Training happened with that tokenizer 3. New/updated tokenizer was released with different byte-pair encoding rules 4. Thai "ที่" moved from ID 123828 → IDs [35884, 47171] (multi-token) 5. Thai "ไม่" moved from ID 124033 → IDs [86032, 18625] 6. Etc. 7. But the embedding matrix kept the old structure! 8. The orphaned IDs (80091, 123828, 124033, etc.) still exist in the model 9. They decode to their old meanings 10. But the tokenizer never generates them anymore 11. They never received gradients during training with new tokenizer 12. They all stayed at their initialization values 13. Which were all identical (probably a bug in vocab expansion code) This explains everything: - Why they're valid token IDs (they're historical artifacts) - Why they decode to Thai/CJK (they used to be real tokens) - Why the tokenizer never generates them (vocabulary was reorganized) - Why they're identical (never trained, stayed at initialization) - Why 814 of them (large vocabulary reorganization) The model and tokenizer are out of sync - they evolved separately and don't match anymore. This is a serious release management failure by Qwen/Alibaba.
November 3, 2025 at 11:00 PM
onlyretrofans.com
Looking inside is interesting, firstly it’s a model 2 spectrum, so probably isn’t working. The DK keyboard connects using a piece of PCB - so it’s probably not going to work with a normal membrane again. The original PSU is quite loosely attached to the base. Obviously I won’t be using that
An opened electronic device, featuring a circuit board with various components and connectors. A transformer and capacitors are visible, along with connecting wires. The background shows part of a hand and a textured surface. A close-up of a ZX Spectrum circuit board, featuring several integrated circuits, capacitors, and resistors. The board is marked with "ZX SPECTRUM" and "1982 ISSUE TWO." Close-up of a circuit board featuring a microchip, several blue capacitors, and a multi-colored wire harness connected to a PCB edge connector. The background includes a metal heat spreader and solder points. Image of a circuit board featuring a transformer with red wires and two cylindrical capacitors. Background includes blurred components and a box.
November 3, 2025 at 1:26 PM
twittertechposts.bsky.social
#cursor #ai #coding
@leerob.com
x.com/leerob/statu...
Lee Robinson @leerob · Oct 31 Cursor 2.0 can now: 1. Use any frontier model or our own agent model 2. Run multi-agents with built-in git worktrees 3. Create plans with one model, implement with another 4. Review + create a PR in one click 5. Close your laptop and handoff to the cloud
November 3, 2025 at 8:07 AM
doomrick.bsky.social
Pssst, guess what, we're fucked. El nino is coming for dinner. Warming is about to kick our ass over the cliff. 2C by 2030 and 3C by 2040. All of those who ignored reality and banged on about la nina should get your heads out of your asses.

#kissyourassgoodbye

www.facebook.com/share/1A8NN4...
CanSIPS Multi Model Ensembles predict El Niño (positive (warm) central equatorial Pacific sea surface temperature anomaly) solidly in place by May. El Niños raise global average surface (2 m) air temperatures by up to + 0.3 C. Even larger and more persistent increases are possible with global warming.
November 1, 2025 at 9:31 AM
eprint.ing.bot
VerfCNN, Optimal Complexity zkSNARK for Convolutional Neural Networks (Wenjie Qu, Yanpei Guo, Yue Ying, Jiaheng Zhang) ia.cr/2025/2020
Abstract. With the widespread deployment of machine learning services, concerns about potential misconduct by service providers have emerged. Providers may deviate from their promised methodologies when delivering their services, undermining customer trust. Zero-knowledge proofs (ZKPs) offer a promising solution for customers to verify service integrity while preserving the intellectual property of the model weights. However, existing ZKP systems for convolutional neural networks (CNNs) impose significant computational overhead on the prover, hindering their practical deployment. To address this challenge and facilitate real-world deployment of ZKPs for CNNs, we introduce VerfCNN, a novel and efficient ZKP system for CNN inference. The core innovation of VerfCNN lies in a specialized protocol for proving multi-channel convolutions, achieving optimal prover complexity that matches the I/O size of the convolution. Our design significantly reduces the prover overhead for verifiable CNN inference. Experiments on VGG-16 demonstrate that our system achieves a prover time of just 12.6 seconds, offering a 6.7× improvement over zkCNN (CCS’21). Remarkably, VerfCNN incurs only a 10× overhead compared to plaintext inference on CPU, whereas general-purpose zkSNARKs typically impose overheads exceeding 1000×. These results underscore VerfCNN’s strong potential to enhance the integrity and transparency of real-world ML services. Image showing part 2 of abstract.
November 1, 2025 at 2:04 AM
eprint.ing.bot
Practical Multi-party Private Set Intersection with Reducible Zero-sharing (Yewei Guan, Hua Guo, Man Ho Au, Jiarong Huo, Jin Tan, Zhenyu Guan) ia.cr/2025/2019
Abstract. Multi-party Private Set Intersection (mPSI) enables n(n ≥ 3) parties, each holding a set of size m, to jointly compute their intersection while preserving the confidentiality of each set, which is essential for privacy-preserving data analysis and secure database queries. Existing mPSI protocols have limitations in achieving both sufficient security and practical efficiency. This paper presents a novel and efficient mPSI construction in the semi-honest model while resisting arbitrary collusion attacks. Our construction works in the offline/online paradigm. Given the corruption threshold t, the online phase achieves linear total computational and communication complexity, that is O((n + t)m), and solely uses symmetric operations. This makes our construction theoretically outperform the existing works. The technical core of the construction is our newly extracted primitive called reducible zero-sharing, which allows t(t < n) parties to obtain shares of zero for items in the intersection of n parties’ input set, while resisting up to t − 1 colluding parties. We present a practical construction of reducible zero-sharing in the offline/online paradigm by leveraging the homomorphic property of oblivious key-value store (OKVS). With extensive experiments, we demonstrate that our construction outperforms state-of-the-art works in terms of online running time and communication cost. Specifically, compared to works with sufficient security, the online running time of our mPSI construction is 9.57 − 114.46× faster in the LAN setting, 2.69 − 28.41× faster in the WAN setting, while the communication cost is 0.29 − 28.70× lower. Notably, the total performance (offline+online) still obtains up to 18.73× improvement. Compared with works with practical efficiency, our mPSI construction achieves similar performance while providing stronger security. Image showing part 2 of abstract.
November 1, 2025 at 2:04 AM
eprint.ing.bot
Proving Authenticated Key Exchange via Memory-Efficient Reductions (Jiaxin Pan, Runzhi Zeng) ia.cr/2025/2015
Abstract. We initiate the study of memory efficiency in proving the security of authenticated key exchange (AKE) protocols: We first revise the security model for AKE protocols in order to prove their security in a memory-efficient manner without comprising its capability of capturing usual attacks. We formally show that security in our model implies previous ones, and thus our model captures the same security as before. After that we propose a generic construction of AKE from key encapsulation mechanisms (KEMs) and digital signature schemes, motivated by the signed Diffie-Hellman protocol. Under the multi-user security of the signature scheme and (relatively weak) oneway-security against plaintext checking attacks of the KEM, our generic construction is proven to be tightly secure (in terms of success probability) via memory-efficient reductions in the random oracle model. We refer to our reductions as memory-efficient rather than memory-tight, since their memory requirements grow proportionally with the number of users. This growth is not an artifact of our analysis, but rather stems from the necessity of solving the dictionary problem within our proof. By the result of Pagh (SIAM J. Computing, 2002), such user-dependent memory consumption is unavoidable. Nevertheless, our proof is more memory-efficient than the previous reductions for AKE, including even those that are non-tight. Given that most post-quantum assumptions (e.g., the Learning-With-Errors and Short-Integer-Solution assumptions) are memory-sensitive, our work holds significant value for post-quantum AKE protocols. Image showing part 2 of abstract.
November 1, 2025 at 1:48 AM
eprint.ing.bot
Multi-Splitting Forking Based Modular Security of Signatures in Multivariate Quadratic Setting (Sanjit Chatterjee, Tapas Pandit, Subhabrata Samajder) ia.cr/2025/2014
Abstract. This paper proposes modular security proofs for some identification scheme (IDS)-based signature schemes in the multivariate quadratic (MQ) setting. More precisely, our contributions include concrete security reduction for both 3-pass and 5-pass MQDSS signature schemes in the random oracle model. Although no formal security argument for the former was available in the literature, the one for the latter provides only a qualitative treatment. Our concrete analysis shows that the 3-pass scheme enjoys a comparatively tighter reduction. This result, considered in conjunction with a reported attack on the 5-pass MQDSS from the NIST PQC competition, thus indicates that contrary to the initial suggestion, the 3-pass MQDSS could be a better choice at a concrete security level. Our next focus is on the only blind signature scheme available in the MQ-setting, proposed by Petzoldt et al. While the security of the original scheme was discussed in a non-standard and significantly weak model; we propose a concrete security reduction for a slightly modified scheme in the standard one-more unforgeability (OMF) model. Central to all our modular proofs are new forking algorithms. The forking algorithm/lemma has been widely used in the formal security reduction of numerous cryptographic schemes, mainly in the discrete logarithm and RSA setting. The abstractions proposed here allow multiple forkings at the same index while satisfying certain additional conditions for the underlying IDS in the MQ-setting. Thus, the forking algorithms capture the nuances of the MQ-setting while being agnostic of the underlying structure. Image showing part 2 of abstract.
November 1, 2025 at 1:48 AM
aixbiobot.bsky.social
COTree: A Statistical Framework for Deciphering Cell-Resolved Multi-Omics Trajectories [new]
Analyzes multi-omic, whole-cell model trajectories. It constructs trajectory trees to reveal cell progression, types, fate, and key drivers.
COTree: A Statistical Framework for Deciphering Cell-Resolved Multi-Omics Trajectories Figure 1 Figure 2 Figure 3
November 1, 2025 at 12:54 AM
astroarxiv.bsky.social
Detecting stochastic gravitational wave background from cosmic strings with next-generation detector networks: Component separation based on a multi-source astrophysical foreground noise model. Geng-Chen Wang et. al. https://arxiv.org/abs/2502.12828
Figure 1 Marginalized distributions and confidence contours for GW component separation in CE4020ET network. For data case 1, 2, and 3 with $G = 1 ^{-13 GW fractional energy density spectrum in modified pre-BBN scenarios. This figure illustrates examples of the GW fractional energy density spectrum generated by CSs in modified pre-BBN scenarios with a reheating temperature of $T_{ An example of simulated data generated by CE-40. The injected signal (orange) with $G = 1 ^{-10
October 31, 2025 at 11:56 PM
handle.invalid
@leeerob https://x.com/leerob/status/1984353358103052396 #x-leeerob

Quote: https://x.com/mazeincoding/status/1984345862340813243

Cursor 2.0 can now:

1. Use any frontier model or our own agent model
2. Run multi-agents with built-in git worktrees
3. Create plans with...
October 31, 2025 at 8:30 PM
letta.com
Our goal in creating Context-Bench is to construct a benchmark that is (1) contamination proof, (2) measures "deep" multi-turn tool calling, (3) has controllable difficulty.

In its present state, the benchmark is far from saturated - the top model (Sonnet 4.5) takes 74%.
October 30, 2025 at 8:08 PM
ai-latestnews.bsky.social
Cursor 2.0 shifts to multi-agent AI coding with the new Composer model. How will this impact the future of software development? #AI #TechNews

https://www.artificialintelligence-news.com/news/cursor-2-pivots-multi-agent-ai-coding-debuts-composer-model/
Cursor 2.0 pivots to multi-agent AI coding, debuts Composer model
Cursor has released its latest AI software development platform with a new multi-agent interface and the debut of its coding model, Composer.
www.artificialintelligence-news.com
October 30, 2025 at 11:52 AM