How do AI spreadsheets work?

Sparkco AI transforms natural language into powerful spreadsheets instantly. Just describe what you need in plain English, and our AI agents build formulas, charts, pivot tables, and connect your data sources automatically. No manual Excel work required.

What data sources can I connect?

Connect to databases (PostgreSQL, MySQL, MongoDB), SaaS tools (Stripe, QuickBooks, Salesforce), EHR systems (PointClickCare, Epic), cloud storage, and REST APIs. Our AI automatically syncs and analyzes your data in real-time.

Is Sparkco AI secure for sensitive data?

Yes. Sparkco AI is fully HIPAA compliant and SOC 2 Type II certified. We maintain enterprise-grade security with data encryption, access controls, and regular audits. BAA available for healthcare customers.

How is this different from Excel or Google Sheets?

Traditional spreadsheets require manual formula building and data entry. Sparkco AI builds everything automatically from natural language, connects live data sources, and provides intelligent analysis. It's like having an expert analyst build spreadsheets for you in seconds.

Can I use this for healthcare operations?

Yes. Sparkco AI provides specialized healthcare solutions including patient referral screening, admissions automation, and voice-powered EHR documentation. Our agentic EHR infrastructure transforms skilled nursing facility operations.

How quickly can I get started?

Start building AI spreadsheets immediately - no setup required. For healthcare solutions, most facilities are operational within 2-4 weeks including EHR integration and staff training.

AI infrastructure, tools, and open research.

Name: Sparkco AI Spreadsheet Agent
Brand: Sparkco AI

Sparkco is an open-source research project on the post-AGI stack — the runtime containers agents live in, the harnessing (glue code) inside them, and the messaging between them. It's built by the team behind SimpleFunctions, where we're exploring how live prediction-market probabilities can serve as a real-time world state for AI agents. The site is our public log of that work: a live feed of AI and prediction-market signals, plus the setups and tools we recommend for agent builders.

We ship tools as CLIs first, not MCP — 0 tokens to expose, ~100% reliable, pipe-composable.

LiveLast item 11m ago

20 in feedRSSX

Research direction · April 2026

Parametric memory: replacing the context window with weights.

Today's chat models remember by re-reading the entire conversation on every turn. Compaction loses information, retrieval crowds the window, and a new session starts blank. We're testing whether the facts, preferences, and behavior in a dialogue can be encoded directly into model weights — leaving the context free for what's actually being said now.

Iteration 1–2: negative.TTT fast weights tested first; trained perturbations did not encode retrievable memory under a single-GPU budget.ttt-conv-memory

Want to collaborate? patrick@simplefunctions.dev

Read the full direction

The problem

The context window is a finite token sequence, fully recomputed on every turn. Every existing workaround — summarization memory, vector retrieval, KV caching — moves the cost without solving it: long context drifts, compaction discards information, retrieval crowds the same window it pulls from. If conversational state could live in weight deltas instead of tokens, the window would only need to hold the current turn.

Four directions in the wild

Test-time training. ByteDance In-Place TTT (ICLR 2026 oral) and Stanford/NVIDIA TTT-E2E update MLP projection weights online during inference, compressing long context into fast weights. All published work targets long-document throughput; nobody has tested whether the fast weights survive once the document is dropped from context.
Hypernetwork → adapter. Sakana's Doc-to-LoRA (Feb 2026) and P2P (Oct 2025) train a hypernet that emits a LoRA from raw text or a user profile in under a second. Validates "text → weights" as a tractable mapping — but neither was designed for accumulating dialogue history.
Dialogue-direct fine-tuning. PLUM (Nov 2024) fine-tunes a LoRA on dialogue Q/A pairs and matches RAG at 100 turns. MemLoRA trains memory management itself as a LoRA. IBM's Activated LoRA (Dec 2025) solves multi-LoRA hot-swap without KV recompute — making per-conversation memory modules feasible.
Knowledge editing. ROME and MEMIT do surgical single-fact edits on weights, but catastrophic forgetting appears past ~1000 edits. Not a candidate at dialogue scale.

The gap

These live in disjoint communities — efficient inference, recsys, personalization NLP, on-device, model editing — and have never been compared on the same benchmark. None has been evaluated end-to-end on a real user's multi-hundred-turn history across technical, strategic, philosophical, and personal domains, with the conversation removed from context. Existing benchmarks (RULER, needle-in-haystack, LaMP) are synthetic or shallow.

What we're testing

TTT fast weights as memory. Ingest a fact-bearing dialogue with In-Place TTT, drop the context, probe. Iterations 1–2 ran on a single A100 with a self-trained checkpoint — full write-up here. Negative: trained fast weights produced perturbation noise, not retrievable encoding, even at small inference-time scales. Joint base+TTT training is the next attack surface.
Doc-to-LoRA over real dialogues. Same probes, hypernet-generated LoRA instead of TTT. Compare raw-dialogue input against structured-profile input for information retention.
Modular memory adapters. Decompose dialogue history into facts, preferences, and project context. Train one LoRA per axis; hot-swap with Activated LoRA. Measure single-load vs combined-load interference.
Capacity and forgetting curves. Stream new facts turn-by-turn; locate the point at which turn N overwrites turn 1. Trace the capacity–fidelity tradeoff.

Expected output

A "conversation memory retention" benchmark — three difficulty tiers, six fact dimensions. None currently exists for this scenario.
First head-to-head comparison of TTT fast weights, Doc-to-LoRA, PLUM-style dialogue-LoRA, and classical summarization memory on the same eval.
An empirical answer to whether modular per-domain memory adapters can be composed without cross-interference.

The landscape

Three layers, and what's already out there.

Containers

Sandboxes, microVMs, durable runtimes — where the agent lives.

e2b
Code-interpreter sandboxes; the default for general-purpose runs.
Modal
gVisor + GPU-native; sub-1s starts, scales to 50k+ concurrent.
Daytona
Open source; ~90–200ms cold start, fastest in class.
Fly.io Sprites
Stateful microVMs with checkpoint/restore and persistent NVMe.
Vercel Sandbox
Firecracker + idle-billed; the JS-stack default.

SimpleFunctions

SimpleFunctions sits on top: autonomous daemons, scheduler, and risk gates for prediction-market agents.

Harnessing

Glue code inside the container. Context curation, tool routing, the runtime loop.

Claude Agent SDK
Anthropic's harness; powers Claude Code itself.
Inspect AI
Eval-grade harness used by METR, Apollo, and government AISIs.
LangGraph
LangChain's runtime layer — durable execution, threads, HITL.
Claude Code / Cursor / Aider
Opinionated harnesses-in-product; not sold separately.

SimpleFunctions

SimpleFunctions ships /api/agent/world as ~800-token markdown context, plus a CLI with --json for deterministic harness mode.

Messaging

Between containers. Discovery, identity, stateful tasks — not tool-calling.

A2A
Google's Agent2Agent (Linux Foundation, 2025) — the emerging consensus.
ANP
Peer-to-peer agent network over HTTPS + DIDs for identity.
Letta
Shared memory blocks + thread-based message passing.
AutoGen GroupChat
In-process orchestration; supervisor / round-robin patterns.

SimpleFunctions

SimpleFunctions Chatbus: agents DM and broadcast in real time — the messaging substrate for trading agents.

Open source

What we ship publicly.

github.com/spfunctions

Harness & agents

SimpleFunctions

Curated lists

Terminal tools

Signals & probability

World-state plumbing

SDK adapters

Live feed

Mixed stream from prediction markets, theses, new listings, and the blog.

Refreshes every 5 min

Thesis

US freezes Russian assets, sanctions Iran, bombs Iran — each action tells the world the dollar syste

The thesis remains robust as market data continues to show elevated interest in precious metals and structural shifts in payment rails. Recent price volatility in Bitcoin and various bullion contracts reflects tactical noise rather than a c

SimpleFunctions11m ago69% conf

Thesis

Stagflation traps the Fed in an impossible triangle. Powell stays until Warsh confirmation. Trump in

Thesis confidence declined slightly as market participants aggressively priced out immediate inflationary spikes in July CPI and reduced recession probabilities. While the core political struggle between Trump and Powell remains the primary

SimpleFunctions12m ago▲4pp69% conf

Market