Adrian (@AdrianBarbir) header
Adrian (@AdrianBarbir) profile picture
San Francisco, CA
Joined November 2014
Photo by Adrian (@AdrianBarbir)Photo by Adrian (@AdrianBarbir)Photo by Adrian (@AdrianBarbir)Photo by Adrian (@AdrianBarbir)Photo by Adrian (@AdrianBarbir)Photo by Adrian (@AdrianBarbir)Photo by Adrian (@AdrianBarbir)Photo by Adrian (@AdrianBarbir)Photo by Adrian (@AdrianBarbir)Photo by Adrian (@AdrianBarbir)
we need 10x more AI deployco’s - especially for the mid-market. no amount of AGI will magically create utopia while I’m excited that physical AI is having its moment, software folks should think twice before over-rotating into a 10x tougher problem space, when there’s untapped greenfield in plain sight
58
22
27
541
68,927
Adrian retweeted
it's becoming clear that continuous data from real companies is the optimal path. you get new workflow data constantly, employees flag what's off, you see what models still get wrong, and that becomes the next tasks and graders.
jevon's implications on rl tasks might be somewhat obvious but worth spelling out as models saturate human work, the surface area for capability & alignment gains starts to shift as the industries transform take finance: a good task (literally 3 months ago) might be take an earnings call for a stock and update an existing Excel DCF with new mgmt assumptions model needs taste to "read between the lines" on expectations vs mgmt guides -> you can't take mgmt teams at face value -> the models are now very good at this finance now moves faster -> pms want live updating dashboards instead of excel workbooks as soon as the mgmt call comes in consequently, the new surface area shifts: 1. can the model distill the earnings call to the 3 data points that move the stock such that the pm can make a decision faster than her competitors? 2. can the model make a dashboard that is interactive and in-line with house firm style? 3. can the model leverage historical traces and memory to understand how the pm has historically traded this name?
3
6
1
70
5,284
At @promptingco, we are now the first platform to track how your products show up across Muse, Grok Bot, and Instinct, all in one place. AI agents are becoming the new front door to shopping. Consumers aren't just asking agents what to buy anymore. They're letting agents buy it for them. See which of your products get recommended, which sources agents cite, and where competitors are winning. Know what it takes to get picked. DM me and I'll show you how your brand stacks up against competitors.
29
21
7
171
19,620
Adrian retweeted
we're so early in enterprise ai adoption enterprises buy "ai" from anthropic and openai who have no idea how to show enterprises what ai can do concretely, labs need playbooks that don't exist most public company ceos can't mention a single task that 2x productivity by ai
if enterprises knew how to use ai, they'd spend 1-5% of revenue on tokens given todays capabilities total enterprise revenue is 200T$, 1% of revenue going ot tokens means $2T in token spend $2T in token spend is 50x more than todays $40B
62
28
13
737
97,241
Adrian retweeted
Agents can be AI, but agency is human. It is every individual’s responsibility to empower our own agency, with the help of tools. But the North Star should always remain human centered.
196
383
74
2,449
192,292
Adrian retweeted
We run large-scale FLUX.2 diffusion workloads on AMD MI300X through the Luminal compiler. Our engineer Reis wrote up how we drive down the cost per image by utilizing cheaper flops.
13
16
5
79
11,574
Excited to announce what we’ve been working on at Argos! We’re building the AI context partner for in-house attorneys to complete the bottom 40% of their work faster and more consistently. We live in the greatest age of productivity with the introduction of LLMs that can do real work, yet getting quality outcomes for real-world tasks has proven to be a huge lift that every industry, particularly legal, struggles with. After spending the last year iterating with design partners and building out the platform, we’re stoked to go out of stealth with this and onboard more in-house legal teams. We’re also expanding our team and hiring engineers in San Francisco! If you want to work on applied AI for one of the fastest growing verticals, please reach out!
8
5
2
20
1,610
What @krandiash is saying below gels with what I’m hearing from other smart folks. I’ve summarized it in the graphic below.
A lot of custom post-training work being done today will be deprecated in favor of smarter models and continual learning 10-100x intelligence with continual learning is around the corner Compute, time and effort to heavily customize models is not a great use of resources while these capabilities are emerging. They will disrupt the assumptions people are building around today. It analogizes to adapting open source BERT variants to build models for language in 2019. LLMs disrupted this by solving the same problem through in-context learning with much less data and effort. 1. The base model will become fundamentally smarter, there are 10-100x improvements to be made in pre-training alone. 2. Continual learning algorithms fundamentally change the interface by which models are updated, how easy they are to maintain and use, and how much effort is required to adapt them to your domain and context. 3. The context and data sources from the domain, and the judgment of what good looks like will of course continue to matter. But even there, the level of data curation and feedback needed will shrink dramatically. Of course, it's not possible for folks to sit around and do nothing, but it's good to understand the challenges the current stack might be up against in the near future.
7
6
1
56
17,343
Agreed with everything Omar said; a few other thoughts: - I think RSI will likely start in the harness layer rather than pre/post-training code: it's the cheapest place for an agent to try out hypotheses and fork + modify its own behavior at scale. See e.g. DGM @jennyzhangzt - At any point in time, a lot of info genuinely belongs outside the model because it's (1) private (2) too specific to be worth memorizing (3) too recent and/or changes too fast. We will always necessarily need a fat harness for that stuff
The answer is deceptively simple imo. There’s a notion of a general-purpose harness (like, say, CoT reasoning, proper compaction, or an RLM) that’s essentially just what *the* model should become. A better “model” in the API will eventually wrap it and train accordingly. And a separate notion of a harness that’s last-mile specification over the underlying bland model. This is irreducible to the extent that your task/context diverges from the mode, regardless of capability. Here, the closer your complexity to the minimum number of bits required, the better.
1
9
1
117
15,226
Working with the @reductoai team has been great! excited to keep scaling w/ them!
The @DoeLabs team first found Reducto through our YC Deal while @AdrianBarbir and @YuhuangOu were building the company. Less than a year later, their three-person team has completed 8M+ agent tasks and grown Doe into an AI agent command center. Loved telling the story of how they went from dental workflows to the platform they’re scaling today, with Reducto powering the document infrastructure underneath!
2
32
2,336
The @DoeLabs team first found Reducto through our YC Deal while @AdrianBarbir and @YuhuangOu were building the company. Less than a year later, their three-person team has completed 8M+ agent tasks and grown Doe into an AI agent command center. Loved telling the story of how they went from dental workflows to the platform they’re scaling today, with Reducto powering the document infrastructure underneath!
7
4
4
37
14,035
this kimi k3 thing is huge, the thesis of having a harness that lets you chooses every model and route between every model continues to prove itself the correct one
1
1
12
911
Adrian retweeted
The age of the generalist agent is upon us. Remember that graph that circulated a while ago Showing % spend with coding at the top? Yeah That graph is about to shift heavily
5
1
1
15
2,100
the labs owning model-harness fit was a psyop
Replying to @matei_zaharia
3) Harnesses make a huge difference in cost-performance. The very simple Pi harness (@badlogicgames) got the same success rate as harnesses from the LLM vendors with Opus and GPT 5.5, but at 2x less cost! Seems to be mainly due to smaller inputs to the LLM.
8
2
1
67
10,980
there is a $100B+ shaped business to be built in the collective intelligence space i suspect slack could win this if they really wanted to however, it requires deeper mastery of capabilities than a typical enterprise saas team there are some promising prototypes on discord
27
8
5
219
39,680
Narrative violation: builders still want to build. AI doesn’t kill headcount. It increases ambition. We’re entering a new era of software factories, venture builders, and tiny teams creating massive companies. The builder era is back.
3
3
12
1,375
If you want fungibility across models, you need to build the harness, context, memory, intelligence and routing in the application layer. This allows you fungibility across models and compute location. Now @matanSF you just need to sell a secure instance of this :). Congrats on this milestone.
3
25
1
350
93,698
Every time a new model is released, I’m amazed by how much it changes AI usage among technical people …and how little it changes AI usage for the average person 99%+ of people probably couldn’t name a single model. There is SO much opportunity at the application layer
44
15
4
223
17,238
A peak into the future
This is exactly right. People are starting to look for cheaper model alternatives and realizing two things at once: open-source models are already very good, and the ability to train and serve them efficiently at scale can change the economics pretty meaningfully. Tokens are still being subsidized, demand is ramping quickly, and the compute crunch is likely to persist. That will push companies toward using the right model for each task instead of defaulting to the most expensive one. We’re still early, but I expect open-weight adoption to accelerate much faster than most people think.
1
1
19
8,644
Routing and post-training open-source models won't only give you more accurate systems but also meaningfully faster and cheaper systems as most companies are currently learning (in addition to giving you more control and privacy). The idea that a "frontier" model (by frontier we mean is slightly more accurate on a few very limited benchmarks) will be better for all domains, all tasks, all setups just doesn't hold up! It's marketing for making you pay more!
We partnered with @FireworksAI_HQ to train open-source models for legal. Here's what we found: 1) Hybrid legal agents can beat frontier models on quality and cost by routing selectively to a frontier advisor. We tested a hybrid setup where GLM 5.1 served as the primary worker, routing tasks to Opus 4.7 as an advisor when needed. GLM invoked Opus sparingly, just 0.83 times per task on average. The hybrid setup beat Opus on both quality and cost: 18% all-pass vs 14%, at $368 vs $954 across the same 100 tasks. 2) Post-training can push open models to frontier-level legal performance. On a 100-task slice of our Legal Agent Benchmark (LAB), SFT moved Kimi 2.6's all-pass rate from 11% to 15%, beating Opus' 14%. But the cost gap was even more striking: $84 vs $954 across the same 100 tasks, or ~11x cheaper. We're excited to continue working with @FireworksAI_HQ on the next generation of open-source legal agents.
30
28
2
274
27,605