itsvalisse

𝐄𝐱𝐩𝐥𝐨𝐫𝐢𝐧𝐠 𝐭𝐡𝐞 𝐀𝐈 𝐟𝐫𝐨𝐧𝐭𝐢𝐞𝐫
Tools, insights & hidden gems
Discovering what’s next

[itsvalisse @elixlune]

Google Research just open-sourced Harness².

The idea: don't retrain the model. Improve everything around it.

The weights stay frozen. The agent edits its own harness:
system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, plugins.

After each task it keeps the lessons for the next one.

One improvement step, same model:
WorkBuddy-Bench 64.2 → 75.7 (authors' results, GLM-5)

Prompts are becoming the new weights.

[Chubby♨️ @kimmonismus]

Google Research has open-sourced Harness², a framework that enables AI agents to improve their own prompts, tools, memory, and workflows. It tests different configurations and carries useful lessons into future tasks while keeping the underlying model unchanged.

An increasing number of releases concerning self-improving models are being published.

[itsvalisse @elixlune]

@Taboobuilds Token / tool budget left — not just cost so far, but remaining allowance. Helps you see when an agent is about to explode the bill before it happens.

[itsvalisse @elixlune]

After every real task the agent doesn’t just keep the lesson.
It automatically spawns 3–5 “ghost” versions of the same task — a missing tool, a stricter constraint, an adversarial user, or a domain shift. On those ghosts it quietly runs a lightweight Propose–Probe–Compose loop in the background. Lessons that prove themselves on the ghosts are pre-loaded into the starting harness of the next real task.
Weights stay frozen. Only the harness evolves. The agent keeps improving even when it’s idle.
No one else has this yet.

[itsvalisse @elixlune]

Google Research just open-sourced Harness².

The idea: don't retrain the model. Improve everything around it.

The weights stay frozen. The agent edits its own harness:
system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, plugins.

After each task it keeps the lessons for the next one.

One improvement step, same model:
WorkBuddy-Bench 64.2 → 75.7 (authors' results, GLM-5)

Prompts are becoming the new weights.

[itsvalisse @elixlune]

@zynkbuilds This is clean. Using Claude Code to actually build working physics models instead of just talking about agents is the right direction. The fact-check catch on the ×8 vs ×8.2 is a nice detail too.

[itsvalisse @elixlune]

@elonmusk USB-C good 👍🏻

[itsvalisse @elixlune]

Solid breakdown. Europe actually giving companies a clear path with the standard is a big deal — most of the industry has been flying blind on compliance. Curious how many will actually use it vs just keep waiting for the US to figure something out.

[itsvalisse @elixlune]

Europe published its first harmonized standard under the AI Act yesterday — EN 18286:2026.
If companies building high-risk AI systems follow its quality management rules, they’ll get a presumption of compliance with the law. No more guessing whether what they’re doing is actually allowed.
Meanwhile in the US, they’re still mostly talking about how to regulate frontier models. Anthropic is even walking around Congress with its own draft bill. No formal rules yet — just discussions.
Europe has moved to specifics. The US is still in “let’s talk about it” mode.

[Elon Musk @elonmusk]

Two important changes to timeline recommendations:

- New AI/SI ranking model that predicts probable time spent much better

- Retrieval model trained to surface posts people read all the way through

User active seconds is growing exponentially and hitting all-time highs almost every day:

[itsvalisse @elixlune]

The real signal isn’t the view count. It’s that people are now mass-cleaning their TVs themselves. Manufacturers sell the “smart” layer as a feature. Users are treating it as something to disable and replace with something light. It works better than waiting for updates.

Useful sites:
• https://t.co/4FKhujONS0 — main viral repo with tested lists for ~50 models
• https://t.co/3ZGchzwH2q — popular clean ad-free launcher for Android/Google TV
• https://t.co/GkDAMoir2D — simple tool to safely disable bloat (no root, fully reversible)

[itsvalisse @elixlune]

One TV cleanup post just pulled 5.2 million views and 21k bookmarks in two days. The repo behind it is already past 4k stars.
It’s not some clever hack. It’s a straightforward open-source script that backs up the TV, disables the bloat with pm disable-user (nothing gets deleted), and drops a 250 KB launcher that actually runs on 1 GB of RAM.
Tested lists exist for Xiaomi, TCL, Sony, Hisense, Shield, Chromecast and a bunch of others.
What’s interesting isn’t the tool itself. It’s that this is going mainstream. People are realizing the “smart” part of a modern TV is mostly just a slow, ad-filled skin that the manufacturer will never really clean up. Instead of waiting for the next software update that makes it worse, they’re just turning the junk off and putting on something light.
While brands are still fighting over who has more dimming zones and longer OS support, a growing number of people are treating the built-in Smart TV as the problem, not the solution. And it’s working better than anything the manufacturers are shipping.

[Gathin (Ø,G) @gathint]

Vibe coded my parents' TV this afternoon!

It was showing a sponsored ad before you could even get to Netflix.

Gave Claude Code the model number, it found the TV on wifi, removed 25 Xiaomi packages and wrote a new launcher.

Looks like an Apple TV now lol

[itsvalisse @elixlune]

@Polymarket LIKE LIKE LIKE ❤️

[itsvalisse @elixlune]

@Qyravex Respekt✊

[itsvalisse @elixlune]

@Amos_FACT Please add it to the chat.

[itsvalisse @elixlune]

@rost_roma LMAOOO 😭😂 gotta keep people on their toes

[itsvalisse @elixlune]

Mike, 38, Boston. $92,000 in sports betting debt.

He was good at one thing: reading odds.

So he stopped betting on games and started betting on lunch.

He built an AI that predicts which Boston corner will be busiest tomorrow: weather, Red Sox schedule, office events, college exams.

One lobster roll truck. It went where the AI said.

3 years later: 9 franchised trucks running his model, $1,500/month each. Debt gone.

Same skill. Better game.

[itsvalisse @elixlune]

@Shileisaboy Hello. 👋

[itsvalisse @elixlune]

@Qyravex Yeah, it does look pretty useful!

[itsvalisse @elixlune]

Google Research just open-sourced Harness².

The idea: don't retrain the model. Improve everything around it.

The weights stay frozen. The agent edits its own harness:
system prompt, guardrails, memory, tool guidance, subagents, skills, scripts, plugins.

After each task it keeps the lessons for the next one.

One improvement step, same model:
WorkBuddy-Bench 64.2 → 75.7 (authors' results, GLM-5)

Prompts are becoming the new weights.

[Chubby♨️ @kimmonismus]

Google Research has open-sourced Harness², a framework that enables AI agents to improve their own prompts, tools, memory, and workflows. It tests different configurations and carries useful lessons into future tasks while keeping the underlying model unchanged.

An increasing number of releases concerning self-improving models are being published.

[itsvalisse @elixlune]

@0xSeinz 99% safe per app sounds reassuring until you do the math. For 25 apps, the probability that nothing breaks is only 77.8% — assuming independence. That’s the real Agent Math.

[itsvalisse @elixlune]

@rost_roma The real test is whether I can explain the logic, not just guess the liar. Drop the puzzle.

[Drenson @rost_roma]

4 AI agents shipped a bad deploy at 3am.
One of them did it. Exactly one is lying.
Took me about a minute. Curious how long it takes you, or your model.

[itsvalisse @elixlune]

@SpXMotivator Models are becoming commodities. The real advantage is in the harness, feedback loops, and the ability to continuously improve.