Solving problems everywhere. All the time.
Joined June 2018
- Tweets529
- Following69
- Followers91
- Likes781
Allwhere Solutions retweeted
Google just open-sourced Harness².
(turn any harness into a self-improving system)
a harness is everything wrapped around the model, like the system prompt, tools, memory, skills, and subagents. today a person designs it once, and the agent runs every new task with that same fixed setup.
Harness² lets the agent refine that setup for each new task on its own, with the model weights left untouched.
on every task, the agent proposes edits to its harness, runs a few candidate versions, and composes a better harness by comparing how those runs behaved.
because it learns from the differences between runs, it needs no correct answer or grader for the task.
that covers a single task. Harness² also carries what it learns from one task into the next.
after each task, it saves the harness it settled on along with the lessons that held up across the runs. the next time it refines a harness, it starts from those notes instead of from scratch.
this means every finished task makes the next round of edits more informed.
Google tested this with Gemini 3.5 Flash and GLM-5 on legal, office, and professional work benchmarks. one round of edits was enough to beat their original harnesses, and giving the agent more rounds pushed scores higher.
those extra rounds can go wide or deep. the agent either tries several candidate harnesses at once or keeps refining a single one step by step, using OpenCode or Codex to carry out the work.
you can find the Harness² repo here: github.com/google-research/h…
I wrote about a similar idea for agents already running in production, where the harness should repair itself after a failure instead of waiting for someone to debug the trace by hand.
that piece walks through how @Cometml's Opik (open-source) closes this loop, going from a failing trace to a proposed fix to a regression test that stops the same failure from coming back. it's quoted below.
Allwhere Solutions retweeted
This is a goldmine for anyone building websites with AI 👀
scrolltide.co is the whole library, not just one prompt. Animated, scroll-driven, 3D.
What it does for you:
Your AI sites stop looking flat.
You stop starting from a blank prompt.
You edit everything and make it yours.
Copy the prompt. Paste into Claude, Cursor or Lovable. Launch live.
All of it for less than a coffee.
Save this 🔖 you'll need it before your next project.
Which one would you build first?
Allwhere Solutions retweeted
🐘 Introducing Postgres Internals in 3D.
🌟 pgrun.dev/pi
Stop reading boring diagrams. Start exploring PostgreSQL.
An interactive 3D map to understand how Postgres actually works.
100% free. Built for developers.
xitter.cf/alxshp/status/21100494…
Introducing PGSync 🐘
Real-time PostgreSQL replication. FREE. OSS
Built in Rust. 🦀
Postgres 16 / 17 / 18.
No extensions. No triggers. No complex setup.
→ snapshot
→ replicate
→ stream changes
→ track WAL
→ recover from crashes
→ resume automatically
→ keep databases in sync
~15 MB RAM during streaming.
Built for developers. Built for AI agents.
Open source.
xitter.cf/alxshp/status/21078790…
Allwhere Solutions retweeted
32 million records searched in 30 milliseconds.
That's absurdly fast.
Search performance is one of those things you have to see to believe.
And this is open source, easy to set up, and supports everything from fuzzy search to vector and semantic search.
Allwhere Solutions retweeted
Today, we’re releasing Bonsai 2 2B, 4B, and 9B.
Based on Qwen3.5, these new ternary models bring stronger reasoning, visual understanding, and tool calling to compact footprints designed for everyday devices:
• 2B: 0.43GB
• 4B: 0.95GB
• 9B: 2.0GB
For this release, we focused on what makes small models useful as local assistants: following instructions, avoiding repetitive loops, calling tools on mobile devices, and giving useful answers without unnecessary output. The 9B model uses roughly 40% fewer tokens on standard instruction-following and everyday math tasks. Solving the Riemann hypothesis can wait.
These models are successors to our first-generation Bonsai 1.7B, 4B, and 8B, and provide a foundation for bringing practical agentic functionality to local applications on personal computers, phones, and wearables.
The Bonsai 2 family now includes four sizes spanning 2B to 27B parameters. Developers can choose the balance of capability, memory use, and responsiveness that fits their application and hardware.
Available today under Apache 2.0.
Allwhere Solutions retweeted
if you're using claude to edit videos, you can give it a library of 5.2M+ audio files for $0
sound effects, ambience, music. openverse has a free api. no key needed.
give it api.openverse.org and ask:
“whenever you need sound effects search Openverse to find the best options and use them for tasteful sound design within my video”
no more using your browser to download random clips or trying to generate the audio itself
Allwhere Solutions retweeted
🚨 This is huge.
Fish Audio just opened the weights of Drama 3, and it's the first open TTS model that sounds close to perfect to me.
Other models can switch emotion mid-sentence.
Drama 3 is the first one I've heard where it actually sounds right.
Three tests against Qwen3-TTS, each on one DGX Spark at home: happy turning sad mid-sentence, a sad whisper, and pure excitement. Listen to where the mood drops.
I think this kills ElevenLabs for anyone who has their own compute.
The catch: it's a 35B model (3B active), a 41.5 GB download, and on the Spark it only just keeps up with real time: 10 s of audio take 9.7 s (Qwen3-TTS 7.6 s).
$52M seed round. 15M+ users. But this is what we're most proud of yet.
Introducing Drama 3: the most emotive and controllable TTS model in the world
And yes, it's open-weight.
You can now control how even a single word should sound, change its tone, and shift your voice mid-sentence.
Uncle Roger was impressed too.
Try it here:
s.fish.audio/rqh0l8
Comment “DRAMA” within the next 48 hours for 1 month free.
If you're an enterprise your first month is on us:
s.fish.audio/qz9r0t
Allwhere Solutions retweeted
you can give claude an entire 3d asset library for $0
2,300+ models, materials and lighting environments from poly haven.free api. no api key. commercial use allowed.
give it api.polyhaven.com/assets and ask:
“find a softbox studio lighting setup, download it and load it into my blender scene”
Allwhere Solutions retweeted
We just open-sourced an entire 3D team for Blender.
Introducing the Scenario plugin.
Use it inside Blender, or let Claude Code run it through MCP.
📁/scenario-plugin
┃
┣ 📁 Concept Artist
┃ ┣ 📁 Concept Art From a Prompt
┃ ┗ 📁 References From Your Viewport
┃
┣ 📁 3D Artist
┃ ┣ 📁 Meshes From Text or a Photo
┃ ┗ 📁 Retexture or Remesh the Selected Mesh
┃
┣ 📁 Animator
┃ ┣ 📁 Rig the Selected Character
┃ ┗ 📁 Motion From Text
┃
┣ 📁 Surfacing Artist
┃ ┣ 📁 PBR Materials Onto Any Slot
┃ ┗ 📁 Maps From a Photo
┃
┣ 📁 Layout Artist
┃ ┣ 📁 Greybox From a Description
┃ ┗ 📁 360 Worlds and Skies
┃
┣ 📁 Camera Operator
┃ ┣ 📁 Camera Path From a Sentence
┃ ┗ 📁 Orbits, Dollies, Cranes and Flyovers
┃
┣ 📁 Render Artist
┃ ┣ 📁 Viewport to Finished Still
┃ ┗ 📁 Playblast to Finished Shot
┃
┗ 📁 Sound Designer
┣ 📁 Music, Voice and SFX
┗ 📁 Dropped on Your Sequencer
Comment "BLENDER" and I'll send you the GitHub repo.
Allwhere Solutions retweeted
native 4K video, from a 1.3B model 🔍
T3-Video retrofits Wan2.1-1.3B with window attention so it generates directly in 3840×2176, no upscaler: zoom in and every bush is still there
▶️ on Spaces hf.co/spaces/hugging-apps/t3…
Allwhere Solutions retweeted
Qwen3.8-27B running at nearly 70 tok/s on a 12GB GPU with a 262K context window sounds like one of those numbers that needs an asterisk.
This time, the asterisk is the engineering.
The setup is Mirai S Qwen3.8-27B, a 2.4-bit compressed version that weighs roughly 11GB. The model has been converted into GGUF form for a custom llama.cpp fork, with the compressed weights preserved rather than being re-quantized. The community conversion is designed specifically around this runtime.
The starting point was already impressive.
On an RTX 3090 limited to 12GB of VRAM, the initial llama.cpp port managed roughly:
82K context
~38 tok/s decode
Then CaryPalmerr took over the runtime work and pushed it much further.
Same GPU.
Same model file.
But now:
262,144-token context
~69.9 tok/s decode
11.7GB peak VRAM
~1,124 tok/s prefill
Q8 KV cache
The really clever part is how the memory is being handled.
Instead of forcing the entire KV cache into VRAM, part stays on the GPU while the rest is kept in pinned system RAM. The runtime also uses MTP speculation across different draft depths.
That combination turns 12GB of VRAM from an obvious limitation into something you can actually design around.
And this isn’t just a theoretical hack.
The Qwen3.8-27B model itself officially supports 262K context, but the normal reference serving configurations are aimed at much larger multi-GPU systems.
Getting anywhere near that context on a single 12GB card is therefore a completely different engineering problem.
Then comes the other part of the experiment.
The Mirai weights use compressed trellis codes, which means you can’t simply modify the weights like a normal checkpoint. The project instead measures the refusal direction on the quantized model and applies the transformation during inference.
The reported result:
64/64 harmful prompts refused → 0/64
78/82 held-out prompts refused → 0/82
Decode: 69.9 → 69.7 tok/s
In other words, the runtime modification barely changed throughput.
There are caveats, though.
Those refusal counts were measured with a regex, and with thinking enabled, 1 of 24 tests still refused. Long-context performance also drops significantly, reaching about 27.9 tok/s around 60K context on the reported RTX 3090 setup.
So I wouldn’t reduce this to “uncensored Qwen running at 70 tok/s.”
The more interesting story is everything underneath it.
A 27B model.
2.4 bits per weight.
12GB VRAM.
262K context.
Pinned RAM for KV cache.
MTP speculative decoding.
~70 tok/s.
That is what happens when the inference engine becomes part of the model deployment strategy instead of just being the thing that loads the model.
Local AI is increasingly becoming an optimization problem around memory movement, caching, speculation and avoiding unnecessary GPU work.
The weights are only half the story.
The runtime is becoming the other half.
Allwhere Solutions retweeted
The people wanted it, here it is.
Full disk search, in milliseconds. 17,000x faster than Finder.
github.com/noahdunnagan/fsea…
Allwhere Solutions retweeted
The agents that run my LinkedIn outreach are currently getting a 55% reply rate on cold outreach and are generating hundreds of signups for me completely on autopilot.
I packaged the exact system into a skill that works for any harness so you can copy it for yourself. superagnt.com/skills/warm-ou…
Heres why this system prints qualified replies:
most outbound starts with a list out of apollo, zoominfo or clay. firmographics. who could buy. that list is cold on arrival because "could buy" and "is thinking about this right now" are two different lists.
so mine starts from linkedin posts my buyers are already reacting to. that also means every person on the list actually uses linkedin, since they just reacted to a post on it. most people don't, so a message to a normal list mostly sits unread.
across the client accounts i've worked on, an apollo-style list usually gets about 1/10th the reply rate of one built this way.
the posts come from three places: buyers posting about the problem themselves (hiring for it, breaking under it, shopping for a fix), a watchlist of my competitors and the creators my buyers follow, and the topic itself.
when a buyer writes the post, the author is the best lead on it. anyone reacting to a competitor's product post is already shopping.
every search is written in the words buyers actually post in, because vendor jargon only finds other vendors.
it pulls everyone who reacted or commented, drops anyone i've already contacted plus competitors and their staff, and scores each person 0 to 100 against my ICP with the reason written on the row. anyone under the bar is gone before i've paid for their email.
this is the part everyone skips.
for every lead that clears the bar, it reads their website, their LinkedIn profile & anything else it can find about them on the internet, works out what agent would actually help their operation, and builds it. a real one, sitting there, with a claim link.
so i am not sending "quick question about your outbound."
i am sending "i built you an agent that does X for your company, want me to send over the link?"
it sequences that into heyreach for linkedin and instantly for email, then runs the inbox.
anyone who has done outbound knows reply speed is most of the game, and also the exact thing you quit doing by week three.
every reply fires a webhook. it re-reads the whole thread, tags it interested, not interested or auto-reply, pings me in slack the second someone's interested, and sends them their claim link. i don't touch it.
and every reply feeds the next run. it writes more searches like the ones that got answers, so week four sources better than week one.
the general rule, and it has nothing to do with my software:
use AI to build a personalized, genuinely useful asset and give it away in the first message.
you're not hard selling anyone. you are handing them a thing. and because it's built off their actual company, nobody can mistake it for a template. reply rates go up for both reasons at once.
it doesn't have to be an agent. the skill builds whatever your business is already good at making:
- sell lead gen? a sample list of 10 to 25 of their own buyers
- sell SEO? the 3 to 5 searches their competitors win and they don't
- build websites? a teardown of their homepage with three fixes
- consult? a one-page plan for the exact problem they commented about
the playbook behind the messages is the best outbound advice on the internet in one place. it's built from 25 full-length videos and 60 posts from operators who do this at scale, plus 3 production systems running this exact play. scraped the videos and posts with superagnt's data MCP btw.
run it once in claude code or codex and look at what it staged. once you like it, it turns the whole thing into a small agent team on a schedule. the sourcing side runs about 8 cents per qualified lead in production.
grab it here, paste the install prompt into your agent and let it cook: superagnt.com/skills/warm-ou…
Allwhere Solutions retweeted
Samsung open-sourced a method that shrinks 13B parameter LLM into less than 1 GB.
It's called "LittleBit"
Instead of storing AI weights as standard numbers, they used latent factorization to crush them down to extreme sub-1-bit levels.
In some configurations, they hit 0.1 bits per weight.
But here is where the architecture gets crazy.
When you compress an AI this much, you can stop doing math.
Instead of forcing the hardware to do heavy floating-point multiplication, LittleBit replaces the core computation with a bitwise XOR operation.
It swaps complex matrix math for basic sign flips.
The results rewrite the rules of model deployment:
• Unlocks a massive 11.6x inference speedup relative to standard FP16 models.
• Radically reduces memory footprint and loading bandwidth.
• Maintains robustness in extreme sub-0.5 bit regimes where previous compression methods catastrophically fail.
This is not a clever optimization.
It is the blueprint for running massive, state-of-the-art AI locally on cheap, resource-constrained devices.
Allwhere Solutions retweeted
The entire RAG industry is about to get cooked.
Researchers developed a new RAG approach that bypasses almost everything traditional RAG depends on.
- No vector DB
- No data embeddings
- No chunking
- No similarity search
It's called PageIndex.
Instead of splitting your documents into chunks and loading them into Pinecone, it creates a tree index that lets the LLM reason through them like a human reading a book.
98.7% on FinanceBench. Outperforms every vector RAG on the leaderboard.
100% free. Open source.
Allwhere Solutions retweeted
AI memory is getting f...cking illegal
10 open-source GitHub projects that stop agents from starting from zero EVERY new session
01 Mem0
▸ github.com/mem0ai/mem0
→ 66K+ stars
02 Hindsight
▸ github.com/vectorize-io/hind…
→ retain → recall → reflect
03 memU
▸ github.com/NevaMind-AI/memU
TURN MEMORY INTO KNOWLEDGE
04 Cognee
▸ github.com/topoteretes/cogne…
→ documents + code + conversations → knowledge graph
05 Graphiti
▸ github.com/getzep/graphiti
→ remembers how facts change over time
06 OpenViking
▸ github.com/volcengine/OpenVi…
MAKE THE AGENT STATEFUL
07 Letta
▸ github.com/letta-ai/letta
→ memory + identity across sessions
08 Letta Code
▸ github.com/letta-ai/letta-co…
REMEMBER ACROSS THE STACK
09 OpenMemory
▸ github.com/mem0ai/openmemory
10 Agent Memory Benchmark
▸ github.com/vectorize-io/agen…
the loop:
experience → remember → connect → retrieve → act → update
3 builds I'd actually test:
coding:
Hindsight → Cognee → Letta Code
personal agent:
Mem0 → Graphiti → Letta
company brain:
Cognee → Graphiti → Hindsight
and this is where the sh...t gets interesting:
bigger context isn't the same as memory
a model can hold 1M tokens and still meet you for the first time every morning
these projects change that
save this before your agent forgets it ⭣
