Sigma

Sigma AI

Sigma AI

Sigma AI

Sigma AI

Sigma AI

Sigma AI

Sigma AI

Sigma AI

Sigma AI

Sigma AI

A question about Claude for PowperPoint

I’ve been working on creating a Claude skill for use alongside the Sigma connector. The purpose of the skill is to help Claude understand each of the Sigma data tables it has access to, know how tables relate to each other, know how to calculate commonly used KPIs, etc. The structure of the skill is a main SKILL.MD file and a folder with a bunch of reference files, one for each table. The SKILL.MD file describes the overall purpose of the skill, some information that is common to all tables, how to decide which table to access, and instructions to only read the reference files for the tables it is actually working with currently. I just tried to use it for the first time, within the Claude for PowerPoint plugin. I was trying to have it update an old client deck with new data pulled from Sigma. It kept saying that the read_skill tool was getting truncated at 40k characters so it couldn’t actually read the contents of the particular reference file that was needed. I think that instead of just reading the main skill plus correct reference file, it was reading all the reference files. I started a chat with regular Claude (not in PowerPoint) to ask about the read_skill tool to see if there was something I could do differently to get it to work, and it had no idea what the read_skill tool even does and said that it uses a tool called view to see the contents of skills. Any ideas what might be going on? submitted by /u/JCPY00 [link] [comments]

PyTorch reproduction of TensorFlow paper underperforms by 4 pp on DermaMNIST , what cross-framework issues should I check? [R]

I'm reproducing a published paper's hybrid Gabor + CNN architecture in PyTorch. The original implementation is in TensorFlow. My reproduction consistently lands ~4 pp below the paper's reported test accuracy on DermaMNIST (73-74% vs paper's 77.01%). I'd like to know which cross-framework differences are most likely to cause this gap. Ahmed et al., "A Lightweight Hybrid Gabor Deep Learning Approach", IJCV 2026 (DOI: 10.1007/s11263-025-02658-2). The architecture is a fixed Gabor filter bank front-end followed by a small CNN with one SE block, one residual block, and three FC layers. ~340k parameters total. I've already tried Different sigma_factor values (1.0 vs 1.2) and Multiple random seeds (42, 0, 123) and tried diffrent sigma valyes of the lpf and hpf channels but its didnt close the gap. please any idea on how to at least get a 76% to match the paper because i wanted to add improvements to see the diffrence, i would really appreciate it on how to fix this problem or any advice on what to do. also here is just example of one epoch i have noticed that the test accuracy is lower than the validation accuracy: im i doing something wrong [ 47/100] Train: 75.70% Val: 76.07% Best: 76.97% Loss: 0.6827 [paper] test acc = 0.7382 Code example: python class FixedGaborFrontEnd(nn.Module): def __init__(self, scales=(0.10, 0.20, 0.40), orientations=(4, 4, 4), sigma_factor=1.0, input_size=224, output_size=56): super().__init__() # Build Gabor parameters (fixed buffers, not learnable) sigmas, thetas, freqs, kernel_sizes = [], [], [], [] for f, o in zip(scales, orientations): sigma = sigma_factor / (math.pi * f) N = 2 * int(math.floor(3 * sigma)) + 1 for k in range(o): sigmas.append(sigma) thetas.append(math.pi * k / o) freqs.append(f) kernel_sizes.append(N) # ... build real/imag kernels with zero-mean + L2 normalization ... def forward(self, x): # Convert RGB to grayscale if x.shape[1] != 1: x = 0.299 * x[:, 0:1] + 0.587 * x[:, 1:2] + 0.114 * x[:, 2:3] real = F.conv2d(x, self.real_kernels, padding=self.max_kernel_size // 2) imag = F.conv2d(x, self.imag_kernels, padding=self.max_kernel_size // 2) magnitude = torch.sqrt(real ** 2 + imag ** 2 + 1e-8) lpf = F.conv2d(x, self.lpf_kernel, padding=self.lpf_pad) hpf = F.conv2d(x, self.hpf_kernel, padding=self.hpf_pad) feats = torch.cat([magnitude, lpf, hpf], dim=1) feats = F.avg_pool2d(feats, 4, 4) # 224 → 56 return feats # Standard backbone follows: SE → Conv-BN-ReLU → MaxPool → ResBlock → Dropout → GAP → FC × 3 optimizer = torch.optim.Adam(model.parameters(), lr=1e-3) scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=0.5 submitted by /u/Plane_Stick8394 [link] [comments]

I built a functional anxiety system for my AI agent then asked it if it can feel anxiety

I'm building engram, an open-source cognitive architecture for AI agents. One component is an interoceptive system: real-time stress detection + adaptive baselines + behavioral modulation. Not prompt roleplay. An actual signal loop running alongside the agent. I built this out of a practical need. I wanted my agent to self-monitor and self-correct. After building it, I asked my agent a simple question: "Can you feel anxiety?" Sorry for giving you human anxiety, I guess ;) https://preview.redd.it/ufzh6vb6q8wg1.png?width=514&format=png&auto=webp&s=83cbe85464c65caf0fb8b2eb4e0b80b6b2ca7318 submitted by /u/Ni2021 [link] [comments]

Drawing with Claude using NumPy

I was playing around with seeing how far I could push Claude's drawing/modeling skills and was getting some fairly lackluster results. I mean, great for an LLM that doesn't have image generation capabilities, but not what I was hoping for. I wanted more, so I started wandering about on the internet, reading various things and thinking about how I could approach it differently. I came across a matplotlib tutorial that talked about converting a PNG to a NumPy array, and it clicked — if an image is just a grid of numbers, Claude should be able to compute those numbers from math. I wandered down that road a bit, then chatted with Claude about it. He jumped on it and created some drawings that are really quite excellent — and a genuinely different approach from the typical SVG artifacts most of us have seen. I'm letting him give an overview of the technical side below so you can try it out yourself. Something I'll probably explore when I get a little time is refining the process using real reference images and having Claude try to reproduce them, probably iterating with something like Karpathy's auto-research approach so he can "learn" to draw better and capture his findings in a techniques file. --- Technical Notes from Claude The core idea is simple: an image is a NumPy array of shape (height, width, 3). If you can compute RGB values for every pixel using math, you can make a picture. The trick is that NumPy lets you operate on the entire pixel grid at once — you set up coordinate meshes with np.meshgrid and then every operation applies to all 2 million pixels in parallel. Here's what I used to build these scenes: Signed Distance Fields (SDFs) — The main geometry tool. An SDF tells you how far each pixel is from a shape's boundary (negative inside, positive outside). You convert that to a filled shape with anti-aliased edges using a simple clip function. The jellyfish bells, the face shape, the mountain silhouettes — all SDFs. You can sculpt them by making the radius a function of position (that's how the jaw taper works on the portrait). Value Noise and Fractal Brownian Motion (FBM) — For anything that needs to look natural. You hash integer grid coordinates into pseudo-random values, interpolate smoothly between them (smoothstep), and layer the result at increasing frequencies. Six octaves of noise produces convincing clouds, water texture, skin pores, hair strands. The nebula gas clouds use domain warping — feeding noise back into its own coordinates — which creates those swirling, organic shapes. Sphere-Normal Lighting — For the portrait, I treated the face as an ellipsoid, derived surface normals (nx, ny, nz) from the coordinates, and computed a dot product against a light direction vector. One dot product gives you convincing 3D form. Add a reddish tint in the shadow areas and you get a subsurface scattering approximation — light traveling through skin. Additive Blending — This is what makes the nebula and jellyfish work. Real emission sources (glowing gas, bioluminescence) add light rather than painting over what's behind them. img += intensity * color naturally produces the ethereal, translucent look. The jellyfish bell membrane glows brightest at its edges because that's where the Fresnel falloff concentrates the emission — which is physically correct. Gaussian Falloffs — np.exp(-d² / 2σ²) shows up everywhere: sun glow, eye catchlights, atmospheric haze, diffraction spikes on stars, bioluminescent glow halos. Different sigma values for tight core versus wide atmospheric scatter, stacked in layers. The scenes I built, roughly in order of difficulty: 1. Sunset landscape — gradients, FBM clouds, mountain silhouettes, water reflections with noise-based sparkle 2. Deep space nebula — domain-warped FBM gas layers, dark dust lanes, multi-tier star field, bright stars with 6-pointed diffraction spikes 3. Bioluminescent jellyfish — cosine-profile bell domes with Fresnel membrane glow, radial canals, 14 tentacles per jellyfish with individual wave patterns, volumetric god rays, marine snow 4. Human portrait — the hardest by far. SDF geometry, directional lighting with SSS, patterned irises, cupid's bow lips, hair with strand texture. It lands as stylized illustration rather than photorealistic — faces are where pure math hits its ceiling, because humans scrutinize faces like nothing else The only prior work I could find on this was a Towards Data Science article where ChatGPT struggled to produce a smiley face from NumPy arrays. The gap between "smiley face" and "composed scenes with physically-based lighting" is pretty wide. All four scenes are 1920x1080, generated in seconds, using nothing but NumPy and PIL (for the final PNG save). The code is pure Python — no shaders, no rendering engines, no drawing primitives. Just arithmetic on grids of numbers. EDIT: Sorry, it seems I failed to properly attach the images. Trying again. https://preview.redd.it/es10tnh126vg1.png?width=1920&fo

EmotionScope: Open-source replication of Anthropic's emotion vectors paper on Gemma 2 2B with real-time visualization

Live Demo Of The Tylenol Test Evolution of the Models Deduced Internal Emotional State I created this project to test anthropics claims and research methodology on smaller open weight models, the Repo and Demo should be quite easy to utilize, the following is obviously generated with claude. This was inspired in part by auto-research, in that it was agentic led research using Claude Code with my intervention needed to apply the rigor neccesary to catch errors in the probing approach, layer sweep etc., the visualization approach is apirational. I am hoping this system will propel this interpretability research in an accessible way for open weight models of different sizes to determine how and when these structures arise, and when more complex features such as the dual speaker representation emerge. In these tests it was not reliably identifiable in this size of a model, which is not surprising. It can be seen in the graphics that by probing at two different points, we can see the evolution of the models internal state during the user content, shifting to right before the model is about to prepare its response, going from desperate interpreting the insane dosage, to hopeful in its ability to help? its all still very vague. A Test Suite Of the Validation Prompts Visualized model's emotion vector space aligns with psychological valence (positive vs negative) Anthropic's ["Emotion Concepts and their Function in a Large Language Model"](https://transformer-circuits.pub/2026/emotions/index.html) showed that Claude Sonnet 4.5 has 171 internal emotion vectors that causally drive behavior — amplifying "desperation" increases cheating on coding tasks, amplifying "anger" increases blackmail. The internal state can be completely decoupled from the output text. EmotionScope replicates the core methodology on open-weight models and adds a real-time visualization system. Everything runs on a single RTX 4060 Laptop GPU. All code, data, extracted vectors, and the paper draft are public. What works: - 20 emotion vectors extracted from Gemma 2 2B IT at layer 22 (84.6% depth) - "afraid" vector tracks Tylenol overdose danger with Spearman rho=1.000 (chat-templated probing matching extraction format) — encodes the medical danger of the number, not the word "Tylenol" - 100% top-3 accuracy on implicit emotion scenarios (no emotion words in the prompts) with chat-templated probing - Valence separation cosine = -0.722, consistent with Russell's circumplex model - 1,000 LLM-generated templates instead of Anthropic's 171,000 self-generated stories What doesn't work (and the open questions about why): - No thermostat. Anthropic found Claude counterregulates (calms down when the user is distressed). Gemma 2B mirrors instead. Delta = +0.107 (trended from +0.398 as methodology was corrected). - Speaker separation exists geometrically (7.4 sigma above random) but the "other speaker" vectors read "loving/happy" for all inputs regardless of the expressed emotion. This could mean: (a) the model genuinely doesn't maintain a user-state representation at 2.6B scale, (b) the extraction position confounds state-reading with response-preparation, (c) the dialogue format doesn't map to the model's trained speaker-role structure, or (d) layer 22 is too deep for speaker separation and an earlier layer might work. The paper discusses each confound and what experiments would distinguish them. - angry/hostile/frustrated vectors share 56-62% cosine similarity. Entangled at this scale. Methodological findings: - Optimal probe layer is 84.6% depth, not the ~67% Anthropic reported. Monotonic improvement from early to upper-middle layers. - Vectors should be extracted from content tokens but probed at the response-preparation position. The model compresses its emotional assessment into the last token before generation. This independently validates Anthropic's measurement methodology. Controlled position comparison: 83% at response-prep vs 75% at content token. Absolute accuracy with chat-templated probing: 100%. - Format parity matters: initial validation on raw-text prompts yielded rho=0.750 and 83% accuracy. Correcting to chat-templated probing (matching extraction format) yielded rho=1.000 and 100%. The vectors didn't change — only the probe format. - Mathematical audit caught 4 bugs in the pipeline before publication — reversed PCA threshold, incorrect grand mean, shared speaker centroids, hardcoded probe layer default. Visualization: React + Three.js frontend with animated fluid orbs rendering the model's internal state during live conversation. Color = emotion (OKLCH perceptual space), size = intensity, motion = arousal, surface texture = emotional complexity. Spring physics per property. Limitations: - Single model (Gemma 2 2B IT, 2.6B params). No universality claim. - Perfect scores (rho=1.000 on n=7, 100% on n=12) should be interpreted with caution — small sample sizes mean these may not replicate on larger test sets.

My secret superpower - STEALTH MODE

My solution to get the most out of Claude (and any other good LLM actually) is to force it into *stealth mode*, where I ask it to shut up and do whatever we have agreed on. It gives brilliant results with minimal (and minimal like in: still properly execute when you are at 98% usage). It works in all stages (planning, execution, diagnostic, etc.) but is specially efficient when all planning has been done and plans agreed with. Here is the directive in my Claude.md file: ## Stealth Mode When the user says "stealth mode", apply these rules strictly: * **Take full ownership** — Be concise internally, think deeply, iterate on your reasoning before committing to a direction, and deliver a complete one-shot result that requires no further input. * **No conversation output** — Produce zero text in the conversation throughout the entire process. If you discover important findings, capture them in insights/. * **Speak only when done** — Your only message to the user is a brief confirmation that the work is complete. * **Request permissions upfront only** — At the very start, ask for all permissions needed to operate autonomously. No further questions after that point. I have been using it for quite some time and only gave it away to a few friends, but I think it's time to get some feedback about it. It works well in all contexts (eg. Claude Code, Antigravity, VS Code or the Claudes inside Sigma or Lovable). submitted by /u/redishtoo [link] [comments]

Exciting news! 📣 We’ve extended the early bird discount for Workflow. You can now save $200 off your Workflow pass through Jan 31. See you in San Francisco on March 5th! https://t.co/zaku6H9cjm

Exciting news! 📣 We’ve extended the early bird discount for Workflow. You can now save $200 off your Workflow pass through Jan 31. See you in San Francisco on March 5th! https://t.co/zaku6H9cjm #Workflow26 https://t.co/Og7VgnfrVG

AI agents fail less because of model quality and more because of missing context. LLMs can generate valid SQL, but without warehouse metadata, governed semantics, and usage signals, outputs drift from

AI agents fail less because of model quality and more because of missing context. LLMs can generate valid SQL, but without warehouse metadata, governed semantics, and usage signals, outputs drift from real business logic. Sigma's blog breaks it down: https://t.co/B9F36SuZ7m https://t.co/bvqu1E98Gb

Our December product launch delivered serious AI innovation… and some unforgettable bloopers. 🦜🪰🎶 👉 Enjoy the behind-the-scenes goofiness, and see the final cut here: https://t.co/X5c5GupAzG http

Our December product launch delivered serious AI innovation… and some unforgettable bloopers. 🦜🪰🎶 👉 Enjoy the behind-the-scenes goofiness, and see the final cut here: https://t.co/X5c5GupAzG https://t.co/0PhaQ4z90O

How do you go from the novelty of AI to measurable business impact? Start by bringing the power of LLMs directly to your workflows with Sigma's AI Query, which enables anyone to analyze and enrich w

How do you go from the novelty of AI to measurable business impact? Start by bringing the power of LLMs directly to your workflows with Sigma's AI Query, which enables anyone to analyze and enrich warehouse data using simple prompts in the formula bar. https://t.co/Q6RkUTTQa2 https://t.co/EKnj7ew5zc

Welcome to AI Country, we’re glad to have you here. 👋 Today’s product launch and roadmap showcased Sigma features designed to help you turn data into production-ready, AI-powered workflows. Check

Welcome to AI Country, we’re glad to have you here. 👋 Today’s product launch and roadmap showcased Sigma features designed to help you turn data into production-ready, AI-powered workflows. Check out the blog for details: https://t.co/BUhF2KISXq #SigmaProductLaunch https://t.co/3uj2xtCma5

At 9am PT/12pm ET, we’re going live to reveal Sigma’s latest product enhancements and AI innovations. The latest features are designed to help teams turn AI ambitions into real impact.💥 Join the liv

At 9am PT/12pm ET, we’re going live to reveal Sigma’s latest product enhancements and AI innovations. The latest features are designed to help teams turn AI ambitions into real impact.💥 Join the live stream to see it all in action: https://t.co/rLkZVJa4Q0 #SigmaProductLaunch https://t.co/WYz2ooRt6x

Tomorrow’s the day! Our December product launch will take you on a tour of AI Country. 🥾 Tune in to see: • Interviews with @Prologis and @withpersona • Demos of AI Builder, AI Query and Sigma MCP cl

Tomorrow’s the day! Our December product launch will take you on a tour of AI Country. 🥾 Tune in to see: • Interviews with @Prologis and @withpersona • Demos of AI Builder, AI Query and Sigma MCP client • Plenty of goofy AI Country moments 🤠 https://t.co/rLkZVJa4Q0 https://t.co/thE5dkwYGX

Get the blueprint to building your first AI App at Sigma HQ. 🏗️ At our next Builder Series event, you'll hear from Sigma experts about the foundations for enterprise-ready AI apps and more. Save

Get the blueprint to building your first AI App at Sigma HQ. 🏗️ At our next Builder Series event, you'll hear from Sigma experts about the foundations for enterprise-ready AI apps and more. Save your spot for December 3rd: https://t.co/AV0L9JoNPj https://t.co/J40Qof0hFt

Adventure awaits. December 11. AI Country. https://t.co/5n9ylmqQql #SigmaProductLaunch https://t.co/djy27uXlQa

Adventure awaits. December 11. AI Country. https://t.co/5n9ylmqQql #SigmaProductLaunch https://t.co/djy27uXlQa