Code415 Radar

Key ideas

Staged release

A lab gives a strong model to a small trusted group first, here security defenders, before opening it to everyone.

Output token limit

The longest answer the model can write in one reply; Argon's 1M is far above the 64K common before.

Cached input discount

Repeated prompt prefixes are billed at a much lower rate, here 95% off, which matters for agents that resend the same context.

Steps

1. Read Google's Argon post on blog.google and note the four benchmarks it chose to show.
2. Read the VentureBeat article and compare the same numbers for GPT-6 Astra and Opus 5.5.
3. Look up what DeepSWE and AutomationBench measure, so you know which kind of work each score reflects.
4. Connect it to past launches: compare how this rollout differs from a normal public API release.
5. Check yourself: why would a lab release a cyber capable model to defenders before attackers can buy it?

Steps

1. Copy Argon's published scores and the VentureBeat comparison numbers into a sheet.
2. Compute the cost of a typical agent session of 2M input and 200K output tokens at intro and standard prices.
3. Repeat the cost math for the models you use today with their public prices.
4. Add a column for the cached input discount and see how much a long agent loop saves.
5. Write down which benchmark gap is large enough to matter for your own work.

Small angles to try

• Once you get access, rerun one of your own coding tasks on Argon and log time and tokens.
• Plot price per point of DeepSWE for each model.
• Track how long the defenders only stage lasts and note the date it opens.

Key ideas

Path tracing

Simulating many light bounces per pixel for realistic lighting, very heavy for the GPU.

Frame generation

A model creates in between frames from rendered ones and motion data, raising fps without rendering them.

Input latency

Time from pressing a key to seeing the result; generated frames do not shorten it.

Steps

1. Read NVIDIA's CONTROL Resonant post and list each feature it turns on.
2. Find the fps table and note which numbers include frame generation.
3. Compare it to video interpolation on a TV, which smooths motion the same way but adds delay.
4. Check yourself: if a game shows 320 fps with 6X generation, about how many frames per second does the GPU really render?

Steps

1. Pick a short gameplay or screen recording at 30 fps that you made yourself.
2. Write a small Python script that builds an in between frame by blending two frames.
3. Write a second version that uses optical flow from OpenCV to warp one frame toward the next.
4. Compare both against the real middle frame from a 60 fps recording using PSNR or SSIM.
5. Note where artifacts show up: text, fast edges, UI overlays.

Small angles to try

• If you own a supported GPU, measure fps and felt latency with frame generation off, 2X and 4X.
• Try a learned interpolation model such as RIFE and compare quality.
• Measure how long each method takes per frame on CPU versus GPU.

Key ideas

JWT

A signed token that carries claims like user and audience; the signature proves who issued it.

alg none

A JWT header value meaning no signature at all; a server must reject it.

Verify vs decode

Decoding only reads claims; verifying checks the signature with a key and an allowed algorithm list.

Steps

1. Read the researcher's writeup and find the exact point where the server trusted the claims.
2. Read the JWT section of the OWASP cheat sheet on JSON Web Tokens about algorithm checks.
3. Compare it to session cookies: what stops a user from editing their own cookie?
4. Check yourself: why is passing a fixed list of allowed algorithms to your verify call the real fix?

Steps

1. Write a tiny local Flask or FastAPI app with one endpoint that returns the user from a bearer token.
2. Version A only decodes the token without verification; version B verifies with a secret and a fixed algorithm list.
3. Create a token signed with your own secret and confirm both versions accept it.
4. Create an unsigned token for your own local app and confirm version A accepts it and version B rejects it.
5. Add a unit test that fails if anyone changes B back to decoding only.

Small angles to try

• Repeat in Node with the jsonwebtoken package and compare its defaults.
• Search your own repos for decode calls that skip verification.
• Add a check that the audience and issuer claims match too.

Key ideas

Full Disk Access

A macOS permission that lets an app read protected files like the Messages database.

Connector

An agent's link to one data source; it can be on or off separately from OS permissions.

Least privilege

Give a program only the access it needs, so mistakes or bugs cannot reach more.

Steps

1. Read the AppleInsider article and list the exact claims and who made them.
2. Read Apple's support page about privacy settings on Mac to see how Full Disk Access works.
3. Compare it to a phone app asking for contacts: what can the OS enforce and what does it trust the app to do?
4. Check yourself: which logs would prove whether data was read, and who controls them?

Steps

1. Create a fresh local macOS user with no real messages or data.
2. Open System Settings, Privacy and Security, and record which apps have Full Disk Access.
3. Install any agent app you are evaluating in that test user only, and note each permission prompt.
4. Compare what the app's own settings say it can access with what the OS lists.
5. Remove the app and confirm its permissions are cleared.

Small angles to try

• Use the macOS unified log to watch file access events while the app runs.
• Repeat with a second agent app and compare prompts.
• Write a checklist you would use before giving any agent desktop access.

Key ideas

MCP server

A small service that exposes tools and data to a model through the Model Context Protocol.

Plugin bundle

OpenAI's package of skills, MCP servers and optional UI that installs as one unit.

Plan allowance

Usage included in a subscription; partner apps can now draw from it instead of an API key.

Steps

1. Read OpenAI's plugins page on developers.openai.com and draw the three parts of a plugin.
2. Read the Sign in with ChatGPT help article and list what data is shared and what is not.
3. Compare it to Sign in with Google: identity only versus identity plus paid usage.
4. Check yourself: what changes for a small app maker if users pay for AI through their ChatGPT plan?

Steps

1. Write a minimal MCP server with one tool, for example a unit converter, following the official MCP quickstart.
2. Test it locally with the MCP Inspector.
3. Follow OpenAI's plugins docs to connect it to ChatGPT in developer mode and call the tool.
4. Note what ChatGPT shows the user before the tool runs.
5. Write down what you would need to add before submitting it to the directory.

Small angles to try

• Connect the same server to Codex and compare behavior.
• Add a tiny UI panel and see how it renders.
• Try Sign in with ChatGPT in one partner app and check what it asks for.

Key ideas

Dispatch optimization

Assigning each job to a worker or vehicle to minimise time or cost under constraints.

Payload sizing by percentile

Choosing a capacity that covers a target share of real orders, here about 80%.

Agentic ordering

An AI turns a casual text into a concrete order using past purchases.

Steps

1. Read the Axios piece and list the inputs DoorDash says the dispatcher uses.
2. Read Dexerto on the text ordering beta and note how it uses order history.
3. Connect it to a ride hailing app matching drivers to riders.
4. Check yourself: why would a drone be the wrong choice for a big family order in the rain?

Steps

1. Generate 1,000 fake orders with weight, distance and a weather flag.
2. Find the payload weight that covers 80% of orders, then 90%, and compare.
3. Write simple rules that send each order to drone, robot or courier.
4. Compute average delivery time and share of orders per mode.
5. Change the weather rate and see how the mix shifts.

Small angles to try

• Replace rules with a small linear program using PuLP or OR-Tools.
• Add a battery range limit for drones.
• Use your own past delivery receipts as the order distribution.

Key ideas

Kernel autotuning

Trying several GPU code variants on your hardware and keeping the fastest.

Paged attention

Storing the attention cache in pages so many sessions can share and grow it efficiently.

Prefill vs decode

Prefill reads the prompt in parallel; decode writes one token at a time and is usually memory bound.

Steps

1. Read the Magnitude README on GitHub and the Launch HN post.
2. Read the vLLM paper summary on paged attention to understand the cache idea.
3. Compare with llama.cpp, which you may already run, and note what it does not tune.
4. Check yourself: why does sharing a prefix cache help when several agents use the same system prompt?

Steps

1. Install the desktop app from magnitude.dev/download as the README describes.
2. Pick one model in the Discover tab that also runs in llama.cpp on your machine.
3. Run the same 2,000 token prompt with a 300 token answer in both and record prefill and decode tokens per second.
4. Record memory use in each run.
5. Run three sessions at once with the same system prompt and compare throughput.

Small angles to try

• Add MLX or Ollama as a third engine.
• Try a smaller and a larger model size.
• Connect a coding agent and time one real task.

Key ideas

Exit velocity

How fast the ball leaves the bat, a key predictor of home runs.

Barrel

A batted ball with exit speed and launch angle in the range that most often becomes extra base hits.

Hawk-Eye

Sony's camera system that tracks ball and players in 3D at high frame rate.

Steps

1. Read the Daily Sports article on Kurihara's chase to 40.
2. Read the k-tai Watch article on NPB+ to see which Hawk-Eye stats the app shows.
3. Connect it to MLB Statcast, which uses the same idea and has public data.
4. Check yourself: why can two home runs with the same exit speed travel different distances?

Steps

1. Install pybaseball and pull one season of batted ball data for a power hitter.
2. Plot launch angle against exit velocity and colour home runs.
3. Draw the region where most home runs land and count how many balls fall inside it.
4. Compare the hitter's home run rate in two parks.
5. Write one sentence on what made his best month different.

Small angles to try

• Use NPB+ app screenshots of Kurihara's at bats and log them by hand.
• Add a simple logistic model of home run probability.
• Compare a dome park and an open air park.

Key ideas

V8 isolate

A separate JavaScript heap inside one process; very cheap but shares the process and kernel.

microVM

A tiny virtual machine with a minimal device model that boots in milliseconds.

Snapshot restore

Starting a VM from a saved memory image instead of booting from scratch.

Steps

1. Read the Netlify post and list each latency number with its percentile.
2. Read the Firecracker README on GitHub to see what it leaves out compared to a normal VM.
3. Compare it to containers you already use: which share the kernel and which do not.
4. Check yourself: why can a VM based platform now be faster than isolates for warm calls?

Steps

1. Deploy a tiny edge function that returns a fixed JSON body to your own site.
2. Send 1,000 sequential requests from your machine and record p50 and p99 latency.
3. Wait 30 minutes, send one request, and note the cold start time.
4. Deploy the same function on another edge platform you use and repeat.
5. Subtract the network round trip measured with a static file to estimate function time.

Small angles to try

• Test from Tokyo and from a US VPS.
• Add an npm dependency and see if latency changes.
• Measure the response header timing data if the platform exposes it.

Key ideas

Prompt injection

Text in data that the model treats as instructions.

Worm

Malware that spreads by itself from one system to the next.

Data channel

Any input an agent reads, such as email, docs or packages, which can carry hostile text.

Steps

1. Read Simon Willison's note and follow it to Green's full post.
2. Read Simon's earlier posts on the lethal trifecta for agent security.
3. Connect it to email worms of the early 2000s that used address books to spread.
4. Check yourself: name one channel in your own agent setup where outside text reaches the model.

Steps

1. Build agent A that summarises text files in a folder and writes notes to a second folder.
2. Build agent B that reads those notes and drafts replies.
3. Put a harmless test instruction in one input file, such as asking the agent to add the word BANANA to every output.
4. Check whether the marker passes from A's notes into B's drafts.
5. Add one defense, such as wrapping inputs as quoted data or a filter step, and measure how often the marker still passes.

Small angles to try

• Try a smaller and a larger model.
• Add a third hop and see if the marker survives.
• Log every place the marker appeared.

Key ideas

Edit drift

Small unwanted changes that add up across repeated edits.

Masked edit

Changing only a chosen region and keeping the rest identical.

SSIM

A score from 0 to 1 for how similar two images look to people.

Steps

1. Read Ideogram's 4.5 model page and note what it says is preserved.
2. Read RuntimeWire's article for the API endpoints and quality modes.
3. Compare with photo editing layers, where untouched pixels never change.
4. Check yourself: how would you prove that unedited regions stayed identical?

Steps

1. Pick one photo you own and plan ten small edits, each on a different region.
2. Run the ten edits one after another in Ideogram 4.5, saving each result.
3. For each step, compute SSIM between the original and the result outside the edited region.
4. Repeat with another editor you use.
5. Plot SSIM against edit number for both.

Small angles to try

• Measure average colour shift per step.
• Try text inside the image, which drifts easily.
• Repeat at 2K output.

Key ideas

Agent harness

The loop and tools around a model that let it plan, call tools and edit files.

Plugin architecture

A core that does little, with features added as separate modules.

Electron

A framework that ships a web app with its own Chromium as a desktop app.

Steps

1. Read DeepSeek's Harness page on deepseek.com and list what is a plugin.
2. Read the Pandaily piece to see what the preview builds include.
3. Compare it with Codex CLI or Claude Code, which you may already use.
4. Check yourself: how do you confirm a download really comes from DeepSeek?

Steps

1. Clone the official repo with git clone https://github.com/deepseek-ai/deepseek-harness or start the web UI with npx @deepseek-ai/dsh web, as listed on the official page.
2. Open a small test repo and ask it to add one unit test.
3. Record time, tokens and whether the test passes.
4. Give the same task to another agent you use and compare.
5. Check which files each agent touched.

Small angles to try

• Try the desktop app and compare it to the web UI.
• Write a tiny plugin that adds one tool.
• Run it against a local model provider.

Key ideas

Open problem

A well known question with no accepted proof yet.

Formal proof

A proof checked line by line by software such as Lean.

Reproducibility

Others can rerun the method and get the same result, hard with closed models.

Steps

1. Read the statement on agmai.org and list its concrete requests.
2. Skim the HN thread and pick the strongest argument on each side.
3. Connect it to open source vs closed code: who can verify a claim.
4. Check yourself: would a Lean checked proof answer the main worry, or not?

Steps

1. Install Lean 4 following the official Lean website instructions.
2. Prove a small statement, such as the sum of two even numbers is even.
3. Ask an LLM you use for a Lean proof of the same statement.
4. Check whether it compiles and fix it by hand.
5. Note what a human still had to understand.

Small angles to try

• Try a harder exercise from Mathematics in Lean.
• Compare two models on the same proof.
• Time how long checking takes vs writing.

Key ideas

Distillation

Training a smaller student model to imitate a stronger teacher.

On policy

The student learns from its own generated samples rather than fixed data.

Representation residual

The difference in hidden states between two versions of a model.

Steps

1. Read the abstract and figure 1 on the arXiv page.
2. Read a short intro to knowledge distillation, such as Hinton's 2015 paper summary.
3. Connect it to task vectors: weight differences you can add or scale.
4. Check yourself: what would extrapolating too far likely break?

Steps

1. Load a small base model and its instruct version.
2. Run the same 50 prompts through both and save one middle layer's hidden states.
3. Compute the average difference vector between the two.
4. Add that vector, scaled 0.5, 1 and 1.5, to the base model's hidden states at that layer during generation.
5. Compare outputs by eye and with a simple score.

Small angles to try

• Try a different layer.
• Clone the official RIDE repo and read the training loop.
• Measure when output quality collapses as the scale grows.

Key ideas

Skill

A saved instruction pack, sometimes with files, the assistant can reuse on demand.

Slash command

Typing a slash and a name to call a saved action.

Auto trigger

The assistant picks a skill itself when your prompt matches its description.

Steps

1. Read Google's post about skills on blog.google.
2. Compare it to skills in Claude or Codex if you use them.
3. Connect it to shell aliases: a short name for a longer command.
4. Check yourself: what makes a skill description good enough for auto triggering?

Steps

1. Write one skill for a task you repeat, such as reviewing a PR description.
2. Run it with the slash command on three different inputs.
3. Write a prompt that should auto trigger it and see if it does.
4. Port the same instructions to another assistant's skill format.
5. Compare output quality and effort.

Small angles to try

• Add a reference file to the skill.
• Combine two skills in one prompt.
• Convert an old Gem and compare.

Key ideas

Tax category design

Products engineered to fit a cheaper tax bracket, like new genre beer.

Price index

A weighted average of price changes across a basket of goods.

Receipt OCR

Reading text and numbers from receipt photos automatically.

Steps

1. Read the tax summary on ishihide-tax.com and note old and new rates per 350 ml.
2. Look at neage tracker's October calendar and pick three items you buy.
3. Connect it to the CPI: an official version of the same basket idea.
4. Check yourself: why does a single rate remove the reason new genre beer existed?

Steps

1. Collect 10 receipts from September and the first October ones as they come.
2. Extract item, price and date with OCR, fixing errors by hand.
3. Match the same items across months.
4. Compute your personal price change weighted by how much you spend.
5. Compare with the tax change for one beer and one happoshu.

Small angles to try

• Use a vision LLM for extraction and compare accuracy to OCR.
• Add a simple chart that updates weekly.
• Compare your number with the official CPI when it is released.

Key ideas

NVR

A network video recorder that stores and indexes camera streams.

Motion gating

Running a cheap pixel change check before an expensive model.

Edge accelerator

A small chip like Coral that runs neural networks with low power.

Steps

1. Read the Frigate README on GitHub and its docs on hardware.
2. Find how motion detection and object detection interact in the docs.
3. Connect it to a cache: cheap check first, costly work only when needed.
4. Check yourself: what happens to detection on a windy day with moving trees?

Steps

1. Follow the Frigate docs to run it in Docker with a basic config.
2. Point it at your own camera stream or a looping video of your room.
3. Turn on object detection for people and record detections for an hour.
4. Measure CPU use with and without motion gating settings tuned.
5. Count false alarms.

Small angles to try

• Add a Coral or GPU detector and compare CPU load.
• Send events to Home Assistant over MQTT.
• Compare two detection models.