Pinned Tweet
Presenting Meridian: a line to connect deterministic compute and language model AI.
From Neural Turing Machines and Differentiable Transformers to The Neural Computer, there’s a rich history of trying to combine traditional deterministic computation with the wildly different ial Intelligence.
I’ve spent the last 4 weeks creating a single neural network that has the combined capabilities of a 4B param language model and a deterministic computation engine based on Web Assembly. It allows the AI deterministic integer computations up to 2^32, control flow (while loops and if statements) and a basic filesystem - all implemented as part of the transformer neural network, no external tool calls.
With this architecture adding fewer than 1 million parameters to an existing 4B param language model I can take it from <20% accuracy on arithmetic with 4-digit numbers to 100% accuracy on 4 digit numbers and 99% accuracy on arithmetic up to 2^32 without adversely affecting the language model’s performance on non-mathematical tasks.
The combined model can precisely execute a range of algorithms including checking number for primeness, finding the GCD of two integers and sorting arrays.
From Neural Turing Machines and Differentiable Transformers to The Neural Computer, there’s a rich history of trying to combine traditional deterministic computation with the wildly different ial Intelligence.
I’ve spent the last 4 weeks creating a single neural network that has the combined capabilities of a 4B param language model and a deterministic computation engine based on Web Assembly. It allows the AI deterministic integer computations up to 2^32, control flow (while loops and if statements) and a basic filesystem - all implemented as part of the transformer neural network, no external tool calls.
With this architecture adding fewer than 1 million parameters to an existing 4B param language model I can take it from <20% accuracy on arithmetic with 4-digit numbers to 100% accuracy on 4 digit numbers and 99% accuracy on arithmetic up to 2^32 without adversely affecting the language model’s performance on non-mathematical tasks.
The combined model can precisely execute a range of algorithms including checking number for primeness, finding the GCD of two integers and sorting arrays.
15
40
205
48,255
Andrew Jefferson retweeted
Introducing AgentWeb: reliable web infrastructure for generalized agents.
47/47 real world tasks on live sites. 1.9s median. $3 total.
Request access: agentweb.us
47/47 real world tasks on live sites. 1.9s median. $3 total.
Request access: agentweb.us
254
97
670
618,338
We’re delivering robot training data to each of our customers at more than 100TB per hour
Sign a deal with us and the next day you have 1 PB of ready-to-go annotated egocentric video in your cluster.
Sign a deal with us and the next day you have 1 PB of ready-to-go annotated egocentric video in your cluster.
As our customers scale to training robot models using millions of hours of our recordings, we needed to build a new kind of data infrastructure.
Terabytes of data delivered in minutes, and the capacity to deliver over a Petabyte per day to each of our customers, to multiple ime. That’s more than 10 Gigabytes a second, delivered into any cloud, wherever our customers need it.
Using our customer portal, customers select precisely the data they want and it starts arriving in minutes at their destination, including AWS, GCP, Cloudflare, Hugging Face and any cloud provider who can handle the ingest speed.
We’ve written about how we worked with @huggingface and @Cloudflare to build this secure, scalable infrastructure. Link in comments.
6
5
44
2,358
Andrew Jefferson retweeted
As our customers scale to training robot models using millions of hours of our recordings, we needed to build a new kind of data infrastructure.
Terabytes of data delivered in minutes, and the capacity to deliver over a Petabyte per day to each of our customers, to multiple ime. That’s more than 10 Gigabytes a second, delivered into any cloud, wherever our customers need it.
Using our customer portal, customers select precisely the data they want and it starts arriving in minutes at their destination, including AWS, GCP, Cloudflare, Hugging Face and any cloud provider who can handle the ingest speed.
We’ve written about how we worked with @huggingface and @Cloudflare to build this secure, scalable infrastructure. Link in comments.
Terabytes of data delivered in minutes, and the capacity to deliver over a Petabyte per day to each of our customers, to multiple ime. That’s more than 10 Gigabytes a second, delivered into any cloud, wherever our customers need it.
Using our customer portal, customers select precisely the data they want and it starts arriving in minutes at their destination, including AWS, GCP, Cloudflare, Hugging Face and any cloud provider who can handle the ingest speed.
We’ve written about how we worked with @huggingface and @Cloudflare to build this secure, scalable infrastructure. Link in comments.
11
19
137
40,570
Andrew Jefferson retweeted
nvidia can ship you an ai factory in 14 days. the uk grid can connect it sometime in the 2030s
NVIDIA INTRODUCES DSX POD AI FACTORIES TO ACCELERATE EUROPE’S INFRASTRUCTURE DEPLOYMENT; DSX POD UNITS REDUCE DEPLOYMENT TIME TO AS LITTLE AS 14 DAYS; EUROPEAN ECOSYSTEM BUILDS WORLD’S MOST EFFICIENT AI FACTORIES WITH NVIDIA DSX
1
1
8
597
Andrew Jefferson retweeted
agents need a new kind of storage infrastructure that the major clouds don't provide yet.
s3 files only supports 1,000 file systems per account, and it takes **five minutes** to make a new file system
archil has customers creating millions of file systems every week, in s
data infrastructure is *really* hard to change, make sure that you're picking the right infrastructure when you start to support your growth
s3 files only supports 1,000 file systems per account, and it takes **five minutes** to make a new file system
archil has customers creating millions of file systems every week, in s
data infrastructure is *really* hard to change, make sure that you're picking the right infrastructure when you start to support your growth
5
4
62
3,602
Andrew Jefferson retweeted
Two new egocentric benchmarks from @microagi evaluate VLM readiness for robotics:
MicroPlanBench and MicroAnnotationBench.
tl;dr: Semantic reasoning ≠ perceptual timing
A lot of work to do.
MicroPlanBench (consequential next-action planning)
When predicting the single valid 8s of egocentric video, SOTA models still fail ~1/3 of the time (GPT-6 Astra: 65.6%, GPT-6.1 Sol: 65.5% at 1/5th cost).
- GPT-6 Astra error breakdown: 60% wrong manipulation, 23% wrong hand assignment, 17% wrong object/destination.
- Across all flags, 42% of failures involve tracking the wrong hand.
microagi.ai/benchmarks/mic…
MicroAnnotationBench (continuous temporal grounding)
Knowing what happens ≠ knowing when
Continuous 60s+ dense manipulation video remains unsolved for generalist VLMs:
- GPT-6 Astra/Sol over-segment (~22 segments/clip vs. 13.8 human mean), collapsing precision (~39%) and strict Boundary F1 (<20%).
- Scaling Test-time compute fails to fix perceptual boundaries: scaling Gemini 4 Argon to high reasoning effort dropped its headline score by 4.9 points while costing 4.7x more
microagi.ai/benchmarks/mic…
Currently, Small specialists match/beat frontier APIs at scale.
@microagi's self-hosted nanoannot-1 outperforms the frontier models (Gemini 4 Argon) across the board:
- 54.8 vs 49.6 overall;
- 32.9 vs 27.3 on strict ± 0.667s Boundary F1
- ~1/10th the cost (~$0.30 vs $3.14/video hr & >300x cheaper than GPT 6 Astra)
Solid work by @ErguzelYag, Ömer Adanir & Team.
MicroPlanBench and MicroAnnotationBench.
tl;dr: Semantic reasoning ≠ perceptual timing
A lot of work to do.
MicroPlanBench (consequential next-action planning)
When predicting the single valid 8s of egocentric video, SOTA models still fail ~1/3 of the time (GPT-6 Astra: 65.6%, GPT-6.1 Sol: 65.5% at 1/5th cost).
- GPT-6 Astra error breakdown: 60% wrong manipulation, 23% wrong hand assignment, 17% wrong object/destination.
- Across all flags, 42% of failures involve tracking the wrong hand.
microagi.ai/benchmarks/mic…
MicroAnnotationBench (continuous temporal grounding)
Knowing what happens ≠ knowing when
Continuous 60s+ dense manipulation video remains unsolved for generalist VLMs:
- GPT-6 Astra/Sol over-segment (~22 segments/clip vs. 13.8 human mean), collapsing precision (~39%) and strict Boundary F1 (<20%).
- Scaling Test-time compute fails to fix perceptual boundaries: scaling Gemini 4 Argon to high reasoning effort dropped its headline score by 4.9 points while costing 4.7x more
microagi.ai/benchmarks/mic…
Currently, Small specialists match/beat frontier APIs at scale.
@microagi's self-hosted nanoannot-1 outperforms the frontier models (Gemini 4 Argon) across the board:
- 54.8 vs 49.6 overall;
- 32.9 vs 27.3 on strict ± 0.667s Boundary F1
- ~1/10th the cost (~$0.30 vs $3.14/video hr & >300x cheaper than GPT 6 Astra)
Solid work by @ErguzelYag, Ömer Adanir & Team.
Today we are releasing two new benchmarks:
MicroPlanBench (next-action prediction): GPT-6 Astra and GPT-6.1 Sol lead. Sol costs 1/5 as much.
MicroAnnotationBench (labels + timing on long real-world clips): our nanoannot-1 beats Gemini 4 Argon, at 1/10 the cost.
14
21
105
7,096
Frontier LLMs are like the stereotype of a genius who struggles to change a lightbulb or the oil in their car.
They have great visual and spatial reasoning but they struggle to control a robot body and get it to perform simple tasks. We used tens of thousands of hours of videos to train a specialised model that’s a great handyman, not a math professor
They have great visual and spatial reasoning but they struggle to control a robot body and get it to perform simple tasks. We used tens of thousands of hours of videos to train a specialised model that’s a great handyman, not a math professor
Today we are releasing two new benchmarks:
MicroPlanBench (next-action prediction): GPT-6 Astra and GPT-6.1 Sol lead. Sol costs 1/5 as much.
MicroAnnotationBench (labels + timing on long real-world clips): our nanoannot-1 beats Gemini 4 Argon, at 1/10 the cost.
3
8
43
7,965
Imagine this, but every sandbox is a robot
the newest frontier is agent swarms. watch us launch 700 sandboxes -- all connected to the same file system -- in 10s, then have them collaborate and build a city on that file system
i used sandboxes here, but if i used serverless tool calling? the total cost of this compute ure is packing agents 40x tighter on each CPU and giving them access to the same shared file system context
0
0
3
574
Andrew Jefferson retweeted
Today, we're announcing Archil's new $10/mo plan for individual developers -- making it more accessible for builders to use the world's most advanced AI infrastructure.
To celebrate, we're hosting our inaugural [super fast] hackathon in a couple of weeks at Casco's offices in n prizes, and will be a great way to see how others on the frontier are building.
RSVP here luma.com/t9hsthtk and we'll see you soon!
To celebrate, we're hosting our inaugural [super fast] hackathon in a couple of weeks at Casco's offices in n prizes, and will be a great way to see how others on the frontier are building.
RSVP here luma.com/t9hsthtk and we'll see you soon!
2
10
47
2,942
Andrew Jefferson retweeted
we're about to witness a rapid reshuffling of the "ai infrastructure" as people realize that the tools required to run a single agent are very different than the tools required to run thousands of agents that need to coordinate
the bedrock of what that future looks like? real, distributed file systems that can be mounted from many locations simultaneously in milliseconds.
for the past year, most agent data has just lived on the local disk of the sandbox (persistent or otherwise). this will not scale as agents work with larger data sets and coordinate more amongst each other, you need *shared storage* which allows multiple sandboxes to access the same data
luckily, we've been building for this future for years at Archil, with the worlds fastest shared file system (approaching local disk) that can also reduce the amount of vcpu required for agents by 40x with serverless tool calling
the bedrock of what that future looks like? real, distributed file systems that can be mounted from many locations simultaneously in milliseconds.
for the past year, most agent data has just lived on the local disk of the sandbox (persistent or otherwise). this will not scale as agents work with larger data sets and coordinate more amongst each other, you need *shared storage* which allows multiple sandboxes to access the same data
luckily, we've been building for this future for years at Archil, with the worlds fastest shared file system (approaching local disk) that can also reduce the amount of vcpu required for agents by 40x with serverless tool calling
Do any of the sandbox providers give me the ability to:
- have N sandboxes running with disks that...
- can be mounted on another sandbox (at least 32 at a time)
- without knowing those mount points before the sandbox has started
- and being able to vary them throughout the lifecycle of that one sandbox?
im basically looking for a way to have a parent sandbox, N children sandboxes (where N varies, and has max X concurrency, which would be lower than max of N), and have those children write to a point on the disk that the parent could read from. N readers 1 writer semi-persistent disks.
15
15
280
19,857
Andrew Jefferson retweeted
coolest complaint any company can get rn
those robots will be out of our place and be deployed soon!
we're hiring!
@EastlondonDev @bercankilic
those robots will be out of our place and be deployed soon!
we're hiring!
@EastlondonDev @bercankilic
3
1
40
239,757
Andrew Jefferson retweeted
My dear friends, I’d like to show you my latest toy project
Introducing Handwired: turn your regular JavaScript code into a neural net!
- no training
- exact outputs vs the JS program
- fully differentiable
All runnable in a browser, without GPU
Introducing Handwired: turn your regular JavaScript code into a neural net!
- no training
- exact outputs vs the JS program
- fully differentiable
All runnable in a browser, without GPU
454
1,017
11,335
1,449,840
Andrew Jefferson retweeted
lots of people have asked: "why would anyone optimize for sandbox startup time if you spend most of your time in the session". this is why.
you only unlock serverless tool calling with sandboxes that start in single-digit milliseconds, with all of your data on them. and, when y run agents with 40x less compute.
check out what it looks like with a real agent trajectory
you only unlock serverless tool calling with sandboxes that start in single-digit milliseconds, with all of your data on them. and, when y run agents with 40x less compute.
check out what it looks like with a real agent trajectory
8
3
76
5,107
