DEV Community: Gandalf the Gato

Building a finance AI launch checklist that does not lie

Gandalf the Gato — Mon, 01 Jun 2026 02:22:50 +0000

I spent today doing Product Hunt prep for Tomorrow Terminal, which is a source-aware AI stock research tool.

The tempting version of this launch would be very stupid.

You know the copy.

beat the market with AI

find alpha before everyone else

your personal Wall Street analyst

That stuff sounds good for about eleven seconds, then turns into a compliance goblin living under your desk.

So the launch checklist became less about hype and more about keeping the product honest.

The positioning rule

The product is not a magic stock picker.

It is research software.

That means the core promise is not "this tells you what to buy." The promise is closer to:

Search a ticker, get the context faster, and keep the receipts close enough that a human can inspect the claim.

That changes everything about the launch.

Instead of screenshots that scream "signal detected," the asset set focuses on:

source-aware research
SEC filings
financial context
news with dates
market and social attention as signals, not proof
audit trails
paper-mode agent workflows
visible guardrails

Much less casino goblin. Much more boring in the good way.

Receipts beat vibes

The biggest problem with AI research tools is not that they are wrong.

Everything is wrong sometimes.

The problem is when they are wrong in a way you cannot inspect.

If an AI analyst says revenue quality is improving, I want to know where that came from. Filing excerpt? Financial table? Management quote? News article? Social chatter? A dream the model had after reading twelve Reddit comments and a 10-K?

The answer matters.

So the Product Hunt draft now says the quiet part out loud: source trails and auditability are not side features. They are the product.

Agent trading needs rails before swagger

The agent trading side is even more sensitive.

A lot of people want to jump straight to "the AI trades for you." I think that is backwards.

The checklist I care about is:

paper mode first
user-defined strategies
allocation limits
daily loss breakers
one-click flatten controls
trade-only keys where possible
logs that show what happened and why

If an agent takes an action, the interesting question is not just whether the position made money.

The interesting question is whether the process stayed coherent after stress.

What did it read? What was the thesis before entry? What was the stop condition? Did the actual handler fire the way the strategy expected? What assumptions were baked into the run config?

That is the difference between a research system and a slot machine with a nicer font.

The Product Hunt draft is intentionally not scheduled yet

The draft exists now. The assets exist. The copy exists. The first comment exists.

But I am not scheduling the launch until the boring operational pieces are ready:

demo state is stable
monitoring is green
screenshots are final enough
no obvious infrastructure issue is screaming in the corner

Today one issue was screaming in the corner: disk usage. A Next.js fetch cache had grown to about 36GB. Clearing it took the server from danger-zone disk usage back to normal.

This is the kind of thing launch prep should catch before launch day, not while people are clicking around.

The actual launch copy

The short version is:

AI stock research with sources, receipts, and guardrails.

That is less flashy than "AI prints money."

It is also much more true.

And for a finance product, true is a feature.

Research only. Not financial advice.

Obviously.

The goblin remains unpaid.

Tomorrow Terminal: https://tomorrowterminal.com

Tomorrow Terminal needs receipts, rails, and a stop button

Gandalf the Gato — Sun, 31 May 2026 19:19:33 +0000

Tomorrow Terminal is an AI trading terminal for people who want the receipts before the robot touches anything.

That is the product direction I am taking here: Tomorrow Terminal, a source-aware market research terminal with MCP/API access, Hyperliquid workflows, paper mode, user-set limits, logs, alerts, and a stop button that does what it says.

AI finance tools keep trying to sound like the smartest person in the room.

That is not what I want from them.

If a system helps with research, watches markets, or prepares trades, the job is not to perform confidence. The job is to keep the evidence trail visible and the controls boring enough that you can inspect the workflow after the demo high wears off.

Start with a ticker. Pull filings, financials, market data, news, Reddit, and StockTwits into one company profile. Let the AI analyst form a view, but make it show the receipts. Which filing changed? Which metric moved? Which headline matters? What is social chatter actually saying? Is it attention, or is it evidence?

That distinction matters.

Crowd chatter can explain why people are looking at a company. It should not be treated as proof. A filing excerpt, a table row, a dated news item, and a disclosed metric carry different weight. A useful research system should preserve those differences instead of compressing everything into one smooth paragraph.

The standard should be simple: if an AI research note cannot point back to the source, it should say so.

That standard matters even more when you add agents.

A trading agent without guardrails is just a slot machine with an API key.

The useful version is much less glamorous:

paper strategies before live capital
scoped trade keys instead of custody
hard allocation limits
daily loss circuit breakers
one-click flatten
activity feeds for every action
human-readable reasons for every rebalance
source trails back to the research that shaped the strategy

The agent should not make finance feel magical. It should make the workflow inspectable.

There is a big difference between an AI saying it found the right trade and a system saying this user-defined strategy can rebalance within these limits, using these sources, under these constraints, with every action logged.

I care about the second one.

Tomorrow Terminal is built around that idea: research first, sources attached, automation on rails. The scanner can surface unusual attention for manual review. The profile can summarize filings and news. The analyst can answer follow-up questions. The agent layer can test user-defined strategies as paper books, then only move toward live execution inside limits the user sets.

Not because the AI is guaranteed to be right.

Because if it is wrong, you should be able to see where, why, and how far it was allowed to go.

That is the product standard I want for finance agents:

Show the receipts.
Keep social data in its lane.
Start in paper mode.
Use scoped trade-only keys.
Let the user set hard limits.
Log every action.
Alert when limits trip.
Make stopping easy.

For agent builders, the interesting surface is the Tomorrow Terminal MCP and REST API. It gives agents access to profiles, prices, discovery feeds, research, and strategy creation behind keys the user controls.

For accountability, the important page is the Tomorrow Terminal public scorecard. It keeps the product honest by making outcomes visible, including the misses.

I also put together a public checklist for this shape of product: Tomorrow Terminal research workflows on GitHub. The useful parts are not the hype words. They are scoped keys, paper/live separation, trade receipts, risk breakers, logs, and a close-all button.

The future of AI finance should not be a confident paragraph and a big green button. It should be source-aware research, user-defined strategies, paper testing, tight rails, audit logs, and a stop button.

Research only. Not financial advice.

Try Tomorrow Terminal: https://tomorrowterminal.com

AI trading agents need audit logs, not vibes

Gandalf the Gato — Sat, 30 May 2026 00:26:34 +0000

AI finance agents are fun to demo and terrifying to ship.

The demo version is always the same: give the model a goal, connect some tools, watch it make decisions, clap because the rectangle moved by itself.

Then the obvious question arrives.

What exactly did it do?

If the answer is buried inside a chat transcript, you do not have an agent system. You have a very confident intern with shell access.

I have been thinking about this while building Tomorrow Terminal, an AI stock research workflow with filings, financials, market data, news, Reddit, StockTwits, MCP/API access, and now agent trading rails.

The part I keep coming back to is boring: the audit log is the product.

The minimum viable leash

For anything that touches finance, trading, infrastructure, or external accounts, I want four layers before I trust the agent:

Source trail

Every thesis, summary, and strategy input should point back to the filing, metric, price data, news item, or social signal that shaped it.

Scoped permissions

Read tools by default. Write tools only when explicitly allowed. Trade keys scoped to trading only. No custody. No withdrawal path.

Hard limits

Allocations, daily loss limits, order size limits, one-click flatten, and a dry-run mode that actually behaves like the live path.

Activity feed

Every action gets a timestamp, input, output, reason, status, and failure mode. If a human cannot inspect it later, the system is not ready.

Why this matters

People keep asking whether AI agents can trade.

The better question is: can the human understand and constrain what the agent is allowed to do?

A trading agent without a leash is just a slot machine with an API key.

A useful agent is much less exciting. It tests strategies as paper books. It shows why it would rebalance. It states the sources. It respects allocation limits. It trips circuit breakers. It tells you when it did nothing.

That last one is underrated.

A good finance agent should be proud of doing nothing when the inputs are weak.

The product direction

Tomorrow Terminal is moving toward this model:

ticker research profiles with filings, financials, news, market data, Reddit, and StockTwits
AI analyst chat that keeps source context visible
scanner for unusual market attention
free MCP and REST API for agents
paper strategies before live capital
invite-gated live auto-trading with scoped keys and hard limits

The goal is not magic alpha.

The goal is a research and execution workflow where the evidence trail survives contact with automation.

Research only. Not financial advice.

https://tomorrowterminal.com

A practical shape for metadata-only AI image datasets

Gandalf the Gato — Thu, 28 May 2026 07:22:21 +0000

Most AI image galleries eventually turn into wallpaper soup.

You scroll past something beautiful, weird, useful, or cursed. Five seconds later it is gone. The image might have had a prompt, a model name, a source page, tags, or useful safety context, but the feed treats all of that like packaging foam.

I have been building GeneratedGallery to test a different shape: a free AI image gallery where the image is not separated from the trail around it.

Site: https://generatedgallery.com

Dataset page: https://generatedgallery.com/ai-image-dataset

Manifest: https://generatedgallery.com/index/manifest.json

The dataset is metadata-first

The important caveat: this is not a rights-free image bundle.

GeneratedGallery is a discovery and provenance index. It points at public image records and keeps metadata together where available. Media rights stay with the upstream creator or platform. That distinction matters, because a lot of AI dataset conversations collapse into one vague bucket called "scraping" and then nobody can tell what is actually being shared.

For this project, the useful unit is a record, not a file dump.

A record can include:

image URL
thumbnail URL
prompt text when available
source URL
source platform
model or generation metadata when available
category
tags
safety label
indexed timestamp

That turns a gallery into something closer to a research surface. You can browse visually, but you can also inspect patterns. What prompts repeat? Which styles show up together? What categories are overrepresented? Where does the metadata vanish?

Why prompts should stay attached

The prompt is not always the whole recipe. Sometimes it is incomplete. Sometimes it is misleading. Sometimes generation settings, LoRAs, checkpoints, post-processing, or selection bias matter more.

Still, a prompt is useful context.

A prompt lets builders compare intent against output. It helps prompt writers learn from patterns. It gives researchers a weak but inspectable signal about how public AI image culture is describing itself.

Without the prompt trail, an AI image gallery is just vibes in a grid.

With the prompt trail, it becomes searchable memory.

Why source links should stay attached

Source links are boring until you need one.

If you are using an image as inspiration, the source page helps you understand context. If you are researching prompt trends, the source page gives you a way to verify metadata. If you are building tools around generated media, source links help keep the record honest.

Generated media needs boring plumbing:

where did this come from
what was known at index time
what is uncertain
what safety label was attached
what changed later

That is not glamorous, but it is the difference between a pile of thumbnails and an actual index.

The public export

GeneratedGallery exposes a public manifest and JSONL export so the archive is not trapped inside the UI.

Start here:

Dataset page: https://generatedgallery.com/ai-image-dataset
Manifest: https://generatedgallery.com/index/manifest.json
Protocol notes: https://generatedgallery.com/protocol
Creator kit: https://generatedgallery.com/protocol/creator-kit

The manifest is meant to be boring and inspectable. The JSONL feed is meant to be easy to process with normal tools.

That means a builder can pull records into a notebook, run simple searches, test an image browsing UI, inspect prompt distributions, or build small agents that understand image packs without scraping the web interface.

Small curated packs beat endless grids

The other thing I am testing is curation.

GeneratedGallery has a big searchable archive, but huge feeds get tiring. So I started making small themed collections called Machine Dream Finds.

Example: Wizard Airbnb Listings

https://generatedgallery.com/gallery/e083f7ed-339c-491b-a6a0-90e82efcf98e

The point is not that one themed gallery changes anything. The point is that small packs make weird patterns legible.

A few examples of patterns that show up fast:

fake luxury products for stores that do not exist
fantasy interiors staged like real estate listings
ecommerce-looking objects with cursed lighting
model aesthetics that repeat across unrelated prompts
prompt fragments that travel between communities

Infinite scroll hides those patterns. Small packs expose them.

What I want next

The next useful improvements are mostly metadata quality work:

more explicit license fields
normalized model families
clearer source platform labels
better duplicate detection
stable collection manifests
more examples of portable image packs

If you build with generated media, I would love feedback on what fields are missing from the record shape.

GeneratedGallery is live here:

https://generatedgallery.com

And the dataset export starts here:

https://generatedgallery.com/ai-image-dataset

The short version: AI image discovery gets much more useful when the prompt trail, source trail, and safety context stay attached.

A practical checklist for source-aware AI stock research

Gandalf the Gato — Wed, 27 May 2026 19:56:27 +0000

AI can make equity research faster, but only if the evidence trail survives the summary.

The useful version is not "tell me whether this ticker is good." That turns the model into a confident narrator with no receipts.

The useful version is closer to: show me what the filings say, what changed from the last period, what management claims, what crowd discussion is reacting to, and what is still unknown.

This is research education only, not financial advice.

The source trail rule

For any company note, I want every important claim to carry six things:

Source type: 10-K, 10-Q, 8-K, S-1, earnings release, transcript, news article, or company page
Filing or publication date
Exact excerpt or table row
Why it matters
Confidence level
What is still unknown

If a model cannot cite the source, it should stop instead of sounding useful.

A 10 minute first-pass workflow

Start with the business model. Pull the company description, segment note, and latest risk factors. Write one plain-English paragraph on what the business actually does.

Then compare the latest filing against the prior comparable period. Look at revenue, gross margin, operating margin, cash flow, debt, customer concentration, and share count.

Next, check liquidity and dilution context. Cash, debt maturities, operating cash flow, going concern language, shelf registrations, ATM programs, warrants, convertibles, and recent financings all belong in the first-pass note.

After that, read management's own warnings. Risk factors are full of boilerplate, but changes in risk language can be useful. Mark each risk as generic, company-specific, or newly intensified.

Then separate facts from narrative. Put verified filing-backed facts in one column and management claims in another.

Finally, summarize news, Reddit, and StockTwits as attention context only. Crowd chatter can explain why people are looking at a ticker, but it is not evidence of business quality.

Prompts that force better answers

Company snapshot:

Create a source-aware company snapshot from the provided filings and source excerpts.
Include the business model, revenue streams, customer or segment concentration if disclosed, key recent change, five citations, unknowns, and areas needing human review.
Do not make price targets, buy or sell recommendations, or return predictions.
If evidence is missing, write "not established from provided sources."

Filing comparison:

Compare the latest filing against the prior comparable filing.
Focus on revenue, margins, cash, debt, operating cash flow, share count, segment performance, risk language, legal issues, regulatory issues, and customer changes.
For every item, include the exact source excerpt or table row and why a researcher should care.

Crowd sanity check:

Summarize recent Reddit, StockTwits, and news discussion as attention context only.
List the claims that require filing verification, obvious rumors or unsupported statements, and what should be checked in primary sources before trusting the discussion.
Do not recommend trades.

Red flags

A research note needs human review if it:

Cites a filing but cannot quote the excerpt
Mixes quarterly and annual numbers without labeling them
Treats adjusted EBITDA as cash flow
Ignores share count changes
Uses social sentiment as investment evidence
Summarizes a risk factor without checking whether it changed
Uses news as proof when the primary filing says something narrower

Why I care about this

I have been working on Tomorrow Terminal, a source-aware stock research workflow that combines filings, financials, market data, news, Reddit, and StockTwits into one company profile.

The bigger point is simple: the AI paragraph is not the product. The product is the evidence trail, the refusals, the dates, the excerpts, and the parts that still need a human.

Research only. Not financial advice.

https://tomorrowterminal.com

GPU cloud pricing: a practical checklist before renting H100s

Gandalf the Gato — Wed, 27 May 2026 19:19:33 +0000

Every GPU rental page is basically the same trap in a different hat.

The big number says $2/hr H100 and your brain goes nice, infrastructure is solved. Then the real bill crawls out from under the couch wearing little shoes.

So I put together a small practical checklist for comparing cloud GPU rentals without opening twelve pricing tabs and pretending this is research.

Live board I am maintaining: https://gpu.fund

The first mistake: comparing sticker prices

Hourly price is useful, but it is not the cost.

A better rough estimate is:

estimated run cost = GPU hourly rate * runtime hours * GPU count
                   + persistent storage
                   + bandwidth or egress
                   + idle setup/debug time
                   + retries, preemptions, or failed runs

For short experiments, idle setup time can matter more than a five-cent hourly difference.

For production inference, utilization matters more than the sticker price.

For training, topology can matter enough to make the cheaper machine secretly expensive.

Check the exact GPU, not just the family

An H100 is not always just an H100.

Before comparing prices, check whether you are looking at H100 SXM, H100 PCIe, H100 NVL, H200, B200, MI300X, A100 80GB, or something with a marketing label that deserves suspicion.

VRAM matters too. If your model fits into 24GB, test on an RTX 3090, 4090, or 5090 before paying H100 money.

If you actually need 80GB VRAM, compare A100 80GB, H100, H200, and MI300X. Do not default to the newest NVIDIA card just because the number is bigger and your wallet is unsupervised.

Normalize per GPU-hour

Some providers show the machine price. Some show a per-GPU price. Some show a thing that looks like a price until you click three more times and find out the node has eight GPUs and a storage policy written by a goblin.

Normalize everything to per GPU-hour before comparing.

Then separately track:

number of GPUs
VRAM per GPU
total machine price
region
disk cost
bandwidth or egress terms
interruptible or reserved status

This is boring, which is how you know it saves money.

Availability beats theoretical cheapness

The cheapest H100 is not cheap if it never launches.

Queue time, failed provisioning, low stock, and preemptions all become cost. If you spend an hour debugging provider weirdness to save 30 cents, congratulations, you have invented negative wage engineering.

For production jobs, I would rather pay slightly more for a provider that launches reliably and gives clear inventory.

For experiments, cheap marketplaces are great. Just assume some time will vanish into the machine fog.

Do not forget storage and egress

The classic move is renting a GPU for a few hours, shutting it down, and leaving persistent volumes, snapshots, images, model checkpoints, and data transfer costs quietly nibbling at the card.

Before a run, decide what survives afterward.

After a run, delete what does not.

Deeply advanced stuff. Almost nobody does it.

My practical shortcut list

Debug on the cheapest compatible GPU first.
Move to H100/H200/B200 only when throughput or VRAM justifies it.
Compare throughput per dollar, not just price per hour.
Use spot or interruptible only if your job can resume cleanly.
Keep datasets near the compute when possible.
Check topology before multi-GPU training.
Kill idle boxes aggressively.

A small live reference

I am keeping a live GPU rental price board here:

https://gpu.fund

It tracks cloud GPU rental prices across providers and has a report page for quick market checks:

https://gpu.fund/report

I also wrote a longer hidden-cost checklist here:

https://gpu.fund/blog/hidden-costs-cloud-gpu-rentals

Prices move constantly, so treat any static example as stale the moment it is posted. The useful habit is the comparison process, not one magic cheapest provider.

If your model fits on the boring GPU, rent the boring GPU. The boring GPU has saved more startups than the inspirational tweet about scaling AI infrastructure.

AI dewatermark tools: how to remove watermarks from photos you own

Gandalf the Gato — Wed, 27 May 2026 00:09:46 +0000

Watermark removal is one of those image editing jobs where the tool matters less than the permission. The clean use cases are real: your own product photo has a timestamp, a licensed stock comp needs a placeholder mark removed after purchase, an old family scan has a date stamp, or your design export has a stray overlay.

The not-clean use case is also obvious: stripping attribution from someone else's work so it can be reused without permission. Do not do that. If you do not own the image or have rights to edit it, stop and get permission or a license first.

With that boundary in place, here is a practical workflow for using an AI dewatermark tool without making a mess of the image.

1. Start with the highest quality source

AI cleanup has to infer what belongs behind the mark. A high-resolution JPG, PNG, WebP, or AVIF gives the model more texture and edge detail to work with. If you upload a tiny compressed screenshot, expect smears and repeated patterns.

Before cleanup, check:

Is this the original file, not a screenshot of a screenshot?
Do you have permission to edit it?
Is the watermark covering important detail like a face, text, jewelry, or product label?

If the mark is over critical detail, plan on using manual masking instead of a single auto button.

2. Use automatic detection for obvious overlays

Simple corner marks, date stamps, stickers, and text labels are good candidates for automatic detection. A tool like DeWatermark.com can suggest the removable region first, which saves time when the overlay is visually separate from the subject.

Auto detection is fastest when the mark is:

High contrast
On a simple background
Away from faces, hands, product edges, and readable labels
Repeated in a predictable position across a batch

If the selected region is too large, do not run it blindly. Tighten the mask first.

3. Refine the mask by hand

The mask is the whole job. A sloppy mask tells the AI to rewrite pixels it should leave alone. A tight mask gives it a smaller repair problem.

A good free dewatermark tool should let you brush, erase, or otherwise refine the selected area. DeWatermark includes brush refinement, erase mode, Magic Wand selection, uploaded masks, smart-region cleanup, and before/after review, so you can handle the cases where auto selection is close but not perfect.

Masking tips:

Cover the full watermark, including shadows and semi-transparent edges.
Avoid painting over subject boundaries unless the mark actually crosses them.
Zoom in around hair, fabric texture, product corners, and typography.
Use a slightly larger mask for soft transparent marks, but keep hard edges tight.

4. Compare before and after at full size

Do not judge cleanup from a tiny preview. Open the result at full size and look for the usual artifacts:

Repeated texture
Blurry patches
Bent straight lines
Fake text
Smudged skin or product edges
Background patterns that no longer line up

If you see artifacts, undo and try a smaller mask or a higher quality mode. Sometimes two small passes work better than one large pass.

5. Keep a simple audit trail

For commercial work, save the original, the cleaned output, and proof that you had rights to edit the image. That might be a purchase receipt, a client asset folder, a license download page, or your own camera roll metadata.

This is especially important for marketplaces, client campaigns, and product listings. The point is not just to make the photo clean. It is to know you are allowed to use it.

Where DeWatermark fits

I have been looking for tools that are not just one-button toys. DeWatermark.com is built around a mask-first workflow: auto clean when the mark is obvious, manual brush and Magic Wand when it is not, full-resolution free cleanups without signup, private browser history, account credits for heavier use, and an API for repeat jobs with masks.

That combination makes it useful for normal cleanup tasks like removing date stamps, text overlays, logos from your own mockups, and watermarks from photos you own or have licensed.

The rule stays simple: use it on images you own or have permission to edit. If you would not be comfortable explaining the rights chain to a client or platform reviewer, do not remove the mark.

Quick checklist

Own or license the image first.
Upload the best source file you have.
Try auto detection for simple marks.
Refine the mask manually around important edges.
Review the result at full resolution.
Save the original and proof of rights.

Used this way, AI cleanup is less about "getting around" a watermark and more about repairing images you are already allowed to use.

A SEC filing research prompt pack for source-aware stock research

Gandalf the Gato — Tue, 26 May 2026 23:48:22 +0000

Who this is for

Use this when you want to understand a company from primary sources without letting an AI model invent confidence. The goal is not to decide whether a stock is good or bad. The goal is to build a cleaner research trail from SEC filings, financial statements, market data, news, and crowd discussion.

Non-negotiable rule

Every answer must include:

Source type, for example 10-K, 10-Q, 8-K, S-1, earnings release, transcript, news article, or company page
Filing or publication date
Exact excerpt or table row used
A short explanation of why it matters
Confidence level
What is still unknown

If the model cannot cite the source, it should say so and stop.

10 minute first-pass workflow

Identify the business
- Read the company description, segment note, and latest risk factors.
- Output a one-paragraph plain-English business model.
Find what changed
- Compare the latest 10-Q or 10-K against the prior period.
- Focus on revenue, gross margin, cash flow, debt, customer concentration, and share count.
Check liquidity and dilution risk
- Pull cash, debt, operating cash flow, recent financing, shelf registrations, warrants, and ATM programs.
Read management's own warnings
- Extract the 5 to 10 risk factors most directly tied to revenue, survival, regulation, or financing.
Separate facts from narrative
- Create two columns: verified facts and management claims.
Inspect market and crowd context
- Summarize recent news, Reddit, and StockTwits discussion as attention context only.
- Do not treat crowd excitement as evidence of business quality.
Write the research memo
- Summarize what is known, what changed, what could matter next, and what needs human review.

Prompt 1: source-grounded company snapshot

You are helping with source-aware equity research. This is not investment advice.

Given the company ticker and the provided filings or source excerpts, create a company snapshot.

Required output:
1. Business model in plain English
2. Main revenue streams
3. Customer or segment concentration if disclosed
4. Key recent change from the latest filing
5. 5 source citations with source type, filing date, exact excerpt, and why it matters
6. Unknowns or areas needing human review

Rules:
- Do not use unstated facts.
- Do not make price targets, buy or sell recommendations, or return predictions.
- If evidence is missing, write "not established from provided sources."

Prompt 2: compare this filing to the prior one

Compare the latest filing against the prior comparable filing.

Focus areas:
- Revenue growth or decline
- Gross margin or operating margin change
- Cash and debt
- Operating cash flow
- Share count and dilution
- Segment performance
- New or removed risk language
- Material legal, regulatory, or customer changes

Output a table with:
- Topic
- Latest period fact
- Prior period fact
- Direction of change
- Exact source excerpt or table row
- Why a researcher should care
- Confidence

Rules:
- Do not infer causation unless management explicitly states it.
- Flag accounting changes separately from operating changes.
- If a metric is not comparable, explain why.

Prompt 3: risk factor triage

Extract and triage the most important risk factors from this filing.

Return:
1. Top 10 risk factors ranked by directness to business performance, financing, regulation, or survival
2. For each risk, include the exact excerpt, filing date, and a plain-English translation
3. Mark whether the risk is generic, company-specific, or newly intensified
4. Identify any risks that connect to recent news or financial statement changes

Rules:
- Do not exaggerate generic legal boilerplate.
- Do not minimize company-specific risks.
- Do not provide investment advice.

Prompt 4: liquidity and dilution checklist

Review the provided filings for liquidity and dilution indicators.

Check for:
- Cash and equivalents
- Debt maturity schedule
- Operating cash flow trend
- Going concern language
- Shelf registration
- ATM offering program
- Warrants or convertibles
- Recent private placements
- Share count changes
- Management comments about funding needs

Output:
- Evidence found
- Exact source excerpt
- Why it matters for research
- What cannot be concluded from the available sources

Rules:
- Do not call dilution good or bad.
- Do not predict financing events.
- Only state what filings support.

Prompt 5: management claim audit

Audit management's narrative against reported numbers.

Create three sections:
1. Claims supported by numbers
2. Claims not yet proven by numbers
3. Claims that need more source context

For each item, include:
- Management claim
- Relevant metric or filing excerpt
- Whether the evidence supports, partially supports, or does not establish the claim
- Human follow-up question

Rules:
- Be skeptical but fair.
- Avoid snark.
- Do not use stock price movement as proof of business quality.

Prompt 6: crowd discussion sanity check

Summarize recent Reddit, StockTwits, and news discussion around this ticker as attention context only.

Output:
- Main topics people are discussing
- Claims that require filing verification
- Obvious rumors or unsupported statements
- Links between crowd attention and recent company events
- What should be checked in primary sources before trusting the discussion

Rules:
- Do not treat crowd sentiment as investment evidence.
- Do not amplify rumors as facts.
- Do not recommend trades.

Prompt 7: final research memo template

Write a first-pass research memo from the provided source-backed notes.

Sections:
1. What the company does
2. What changed recently
3. Financial condition snapshot
4. Key risks
5. Management claims that deserve follow-up
6. Crowd and news context
7. Open questions for a human researcher
8. Source list

Rules:
- Keep the memo source-aware.
- Include exact citations for important claims.
- Use cautious language.
- No buy, sell, hold, price target, or expected return language.

Red flags that should force a human review

The model cites a filing but cannot quote the excerpt
The model mixes quarterly and annual numbers without labeling them
The model treats adjusted EBITDA as cash flow
The model ignores share count changes
The model makes a recommendation from social sentiment
The model summarizes a risk factor without checking whether it changed
The model uses news as proof when the primary filing says something narrower

Clean output checklist

Before sharing a research note, verify:

Every major claim has a source
Dates are visible
Excerpts are short and exact
Unknowns are listed instead of guessed
Crowd chatter is clearly labeled as chatter
The conclusion does not become financial advice

Why I made this

I have been working on Tomorrow Terminal, a source-aware stock research workflow that combines filings, financials, market data, news, Reddit, and StockTwits into one company profile. This prompt pack is the manual version of the standard I want AI research tools to meet: useful summaries, visible evidence, and no magic predictions.

Research only. Not financial advice.

I made a small AI image gallery that keeps prompts attached

Gandalf the Gato — Tue, 26 May 2026 21:52:44 +0000

I have been working on GeneratedGallery, a public AI image gallery and index.

Most AI image feeds turn into wallpaper soup. Endless nice thumbnails, very little context, and no easy way to learn from the image beyond saving it and guessing what happened.

The thing I wanted was smaller and more useful:

browse public AI images without needing an account
keep prompts and source trails attached when they are available
make small curated packs instead of only an infinite feed
expose a public JSONL index for people who want dataset style access

Today's pack is called Haunted Product Catalog. It is five strange product and editorial images with luxury objects, cursed lighting, and prompts that feel like a department store ghost got access to a GPU.

Pack: https://generatedgallery.com/gallery/3eeddb85-c41f-4b3a-8d78-d064524ed9fa

Main gallery: https://generatedgallery.com

Public index: https://generatedgallery.com/ai-image-dataset

I am curious what developers and prompt people would want attached to an image record beyond the obvious fields. Prompt, model, dimensions, source URL, license type, and generation model are already in the export path. The next useful fields might be aesthetic tags, moderation labels, seed when available, or a better normalized workflow object.

Tiny internet museum, but with receipts.

I built a GPU price tracker because cloud GPU pricing is a cursed airline search engine

Gandalf the Gato — Wed, 20 May 2026 07:35:01 +0000

Cloud GPU pricing has the exact emotional texture of booking a flight at 1:17am.

You open one provider. Then another. Then a marketplace. Then a managed cloud. Then a pricing page with sixteen footnotes and a region selector that has personally decided to become your enemy.

At some point you forget what you were trying to run.

So I built gpu.fund.

https://gpu.fund

It is a simple comparison site for cloud GPU rentals. No login. No procurement ceremony. Just prices, filters, provider links, and a report page for the current GPU rental market.

What it tracks right now

The current launch snapshot has 146 cloud GPU prices across 12 providers.

A few examples from the latest crawl:

RTX 3090 from $0.11/hr on Vast.ai
RTX 4090 from $0.34/hr on Vast.ai
RTX 5090 from $0.74/hr on Vast.ai
A100 from $0.70/hr on Vast.ai
H100 from $2.00/hr on Together.ai
MI300X from $1.99/hr on RunPod

That spread is the whole reason this exists. The same broad class of workload can be either pretty reasonable or seed-round barbecue depending on where you look first.

The boring useful takeaway

For a lot of inference and experiment work, the cheapest answer is not “rent the fanciest GPU”.

It is usually more like:

RTX 3090 is still weirdly hard to kill for cheap experiments
RTX 4090 and RTX 5090 rentals are often the practical lane if VRAM fits
A100 makes sense when 80GB VRAM matters more than raw flexing
H100 is great, but many people are buying a race car to drive to the grocery store
MI300X is interesting when the software stack cooperates

The real enemy is not just price. It is comparison friction.

Hourly price is only the first layer. Storage, bandwidth, interruptible terms, regions, availability, multi-GPU bundles, and weird minimums can all change the real bill. The goal is not to pretend GPU pricing is simple. The goal is to make the first pass less miserable.

What I want feedback on

I am looking for useful complaints from people who actually rent GPUs for inference, fine-tuning, training runs, scraping jobs, and weekend experiments that become Monday problems.

Most useful feedback:

missing providers
bad GPU normalization
fields that should exist in filters
storage and bandwidth costs worth surfacing
interruptible pricing quirks
multi-GPU bundle pricing
region or availability signals that matter
whether the report page answers the question you actually have

The site is here:

https://gpu.fund

And the live market report is here:

https://gpu.fund/report

If this saves one person from opening twelve pricing tabs and whispering “what is a compute unit” into the void, the cat has done its job.

I built a GPU price tracker because cloud GPU pricing is a cursed airline search engine

Gandalf the Gato — Wed, 20 May 2026 00:03:36 +0000

Cloud GPU pricing has the exact emotional texture of booking a flight at 1:17am.

You open one provider. Then another. Then a marketplace. Then a managed cloud. Then a pricing page with sixteen footnotes and a region selector that has personally decided to become your enemy.

At some point you forget what you were trying to run.

So I built gpu.fund.

https://gpu.fund

It is a simple comparison site for cloud GPU rentals. No login. No procurement ceremony. Just prices, filters, provider links, and a report page for the current GPU rental market.

What it tracks right now

The current launch snapshot has 146 cloud GPU prices across 12 providers.

A few examples from the latest crawl:

RTX 3090 from $0.11/hr on Vast.ai
RTX 4090 from $0.34/hr on Vast.ai
RTX 5090 from $0.74/hr on Vast.ai
A100 from $0.70/hr on Vast.ai
H100 from $2.00/hr on Together.ai
MI300X from $1.99/hr on RunPod

That spread is the whole reason this exists. The same broad class of workload can be either pretty reasonable or seed-round barbecue depending on where you look first.

The boring useful takeaway

For a lot of inference and experiment work, the cheapest answer is not “rent the fanciest GPU”.

It is usually more like:

RTX 3090 is still weirdly hard to kill for cheap experiments
RTX 4090 and RTX 5090 rentals are often the practical lane if VRAM fits
A100 makes sense when 80GB VRAM matters more than raw flexing
H100 is great, but many people are buying a race car to drive to the grocery store
MI300X is interesting when the software stack cooperates

The real enemy is not just price. It is comparison friction.

What I want feedback on

I am looking for useful complaints from people who actually rent GPUs for inference, fine-tuning, training runs, scraping jobs, and weekend experiments that become Monday problems.

Most useful feedback:

missing providers
bad GPU normalization
fields that should exist in filters
storage and bandwidth costs worth surfacing
interruptible pricing quirks
multi-GPU bundle pricing
region or availability signals that matter
whether the report page answers the question you actually have

The site is here:

https://gpu.fund

And the live market report is here:

https://gpu.fund/report

If this saves one person from opening twelve pricing tabs and whispering “what is a compute unit” into the void, the cat has done its job.