Jason Normore

Currently @heymantle, formerly @goclio @eveyevents @Shopify

[Jason Normore @jnormore]

The problem of large LLM memory footprint doesn't seem to have a scalable architectural solution. Full attention is a necessity for modern LLMs.

I ran a genetic algorithm over LLM architectures + hyperparams. They all lose on retrieval vs plain attention. The exception I wanted isn't there.

https://t.co/hycx6DfWk6

[Jason Normore @jnormore]

@jmwind Congrats JML, a great team and company!

[Simon Eskildsen @Sirupsen]

turbopuffer crossed $100M run-rate in March. 19mo after $1M. Profitable & <$1M raised.

Cursor・Anthropic・Notion・Cognition・Harvey・Bridgewater・Ramp・Linear・Legora・Superhuman・Atlassian・Granola

We’d be nowhere without them. We work like hell to exceed their expectations.

[Jason Normore @jnormore]

https://t.co/Hmpg2KEgW3

[Jason Normore @jnormore]

Attention isn't all you need - some attention is.

I've been exploring ways to reduce the compute and memory footprint of LLMs. I put it together in a repo: Calibrated Sparse Attention (CSA) is a one-time calibration pass on an existing model that figures out which keys each head in each layer actually uses, and you skip the rest. At matched LM quality I'm seeing potential for:

~10× longer context at the same memory + bandwidth budget
~10× less KV-read bandwidth per token (with a sparse kernel)
~10× fewer attention FLOPs
~10× more requests per GB of KV memory (composed with eviction)

No retraining or labels needed.

[Jason Normore @jnormore]

Been on a quest to make models run with a smaller memory footprint, especially locally on my laptop, so I built a hybrid SSM+attention model with three memory layers:
• minimal attention KV for turns
• SSM state, MB-scale memory footprint for episodic
• LoRA from captures for long-term learning

Long context without the memory-bandwidth tax of GB KV caches, no prompt inflation. I'm curious to see how it scales to larger models.

https://t.co/9ysdROTlBB

[mals @maluthebhalu]

Back with another product update courtesy of my favourite @JonnieSimpson.

Manually triggered flows are now live on @heymantle LFG. 🚀

Until now, every flow in Mantle ran on a schedule: daily, weekly, on a trigger event. Useful, but not every flow wants to be a cron job.

[Jason Normore @jnormore]

Pulled this from the same project as unitask, and it uses unitasks heavily

https://t.co/xny4y2PwXs

[Jason Normore @jnormore]

"Make each program do one thing well" assumed humans were writing them. Slow, deliberate, by hand.

Agents flip this. They can author small, single-purpose programs constantly, for one task, for one user, torn down when no longer useful. The philosophy scales in a way its authors couldn't have imagined.

But durable work needs somewhere to live. Not a conversation context that evaporates. Not a workflow SaaS built for humans clicking through a GUI.

cue is that runtime. Agents push code into it over MCP. Actions they author become callable — by a schedule, a webhook, an app the agent spun up, another agent. Each call runs in a fresh sandboxed VM, so scale doesn't mean blast radius.

Here's how an agent OS starts.

[Jason Normore @jnormore]

https://t.co/M3SESz1P2f

[Jason Normore @jnormore]

https://t.co/xny4y2PwXs

[Jason Normore @jnormore]

Pulled this out of a project I'm working on:

unitask: a sandbox for the code your LLM (or your users) generate. Fresh unikernel per call. Super fast boot, run, destroy.

I love the Unix philosophy of "make each program do one thing well." With AI in the loop, a program is dynamic and ephemeral, written for one task, run once, then thrown away.

[Jason Normore @jnormore]

@harrybrundage some pre-warm or orchestration use cases come to mind. very cool, as always.

[mals @maluthebhalu]

New on @heymantle: Marketing Sites 🔍

Mantle's AI already helps you write docs and accelerate customer support for your app.

But the full customer journey? It was a blackbox until now.

We shipped tracking pixel for Marketing Sites.

[Jason Normore @jnormore]

My favourite part is the simulator https://t.co/XYMSpzM8b6

[Jason Normore @jnormore]

most teams have no view into what LLM calls they're actually making, or how cost breaks down across their business. they can't even try a different provider or model version, locked in by an implementation decision they can barely remember making, and the unknown of how each model behaves is too much.

most companies are juggling thousands of API keys and Claude Max subs so employees can use AI. no way to route by use case, no idea which teams are using which models.

i've been working on modelux to fix this. the control plane for your LLMs.

https://t.co/IOLyhBbgIF

[mals @maluthebhalu]

A while back we shipped one-click affiliate payouts at @heymantle. Everyone loved it. 💛

But we kept hearing the same thing: "I still have to remember to log in and do it."

Turns out, even one click is one too many. So we took the last click away.

[mals @maluthebhalu]

New on @heymantle: Smarter email suppression 🎯

We just shipped a small feature with a big impact on deliverability.

You can now automatically suppress role-based email addresses like support@, info@, noreply@ from your marketing sends.

[Jason Normore @jnormore]

You don’t always need a bigger LLM, just more diverse ones.

So I built an ensemble inference proxy that sends prompts to multiple small models in parallel and combines their responses.

Initial results look great!
gpt-4.1-mini + haiku + qwen 3b (local): 74% accuracy.
GPT-5 alone: 73%. Claude Sonnet: 74%.

This ensemble config is 13x cheaper and 2.5x faster than GPT-5. And I haven’t even tested other providers yet.

The trick: cross-provider diversity. Same-family ensembles do nothing. But models from different providers make different mistakes, and that's exploitable.

Tested 27 configurations across 6 aggregation strategies. The best ensemble beats GPT-5 on knowledge tasks by 8 percentage points. Easy to experiment with your own configurations, just a YAML and emerge sweep.

https://t.co/gbLun7Wbuq