Pinned Tweet
we @decasonic continue to be actively deploying capital in early stage AI (software and physical) and crypto companies.
we are a team of investor operators. I have developed my team’s AI OS - and post trained models. if you are actively aiming to crush the frontier, and are looking for aligned partners - my inbox is open.
we are a team of investor operators. I have developed my team’s AI OS - and post trained models. if you are actively aiming to crush the frontier, and are looking for aligned partners - my inbox is open.
3
1
17
962
part of the ease of software creation is the development of experiences at scale. often times what this leads to is the development of (1) one-time software, non-reusable and (2) standalone software.
assistant-first platforms similar to @Muse and @wabi could provide the ability for users to more efficiently utilize the software being created through:
1/ reduce tech debt from the need to maintain the many applications at scale. in theory, now it's one prompt to muse (or done proactively in the background)
2/ mitigate against hosting considerations (one click deploy. no need to figure out cloud considerations - similar to Replit and Lovable). of course this also includes security considerations
3/ enable remixing of software that could allow individuals to build and leverage on top of one-time software (system-level network vs individual standalone software). imo this is the crown jewel that leads to the ability to build network effects through multiplayer experiences.
assistant-first platforms similar to @Muse and @wabi could provide the ability for users to more efficiently utilize the software being created through:
1/ reduce tech debt from the need to maintain the many applications at scale. in theory, now it's one prompt to muse (or done proactively in the background)
2/ mitigate against hosting considerations (one click deploy. no need to figure out cloud considerations - similar to Replit and Lovable). of course this also includes security considerations
3/ enable remixing of software that could allow individuals to build and leverage on top of one-time software (system-level network vs individual standalone software). imo this is the crown jewel that leads to the ability to build network effects through multiplayer experiences.
Design the Muse app that’s perfect for you.
Build spaces for whatever you want, pick what lives in your tabs, and even set when they show up.
Here are a few we’ve made so far ↓
1
2
3
169
Abdul retweeted
What the hell?!! I can't believe how far we've come. Atlas is so good. Look at this fucking splat!! Check out behind the DJ table. Totally consistent novel view. No ghosting. This is entirely generated from a single image.
I'm gonna turn this into such a sick VR experience.
I'm gonna turn this into such a sick VR experience.
25
21
251
20,345
there's naturally a significant amount of discussion today about the future of AI investing and spending given the state of the midterms.
the main positioning of how we think about the state of AI post midterms is primarily centered around the migration of value accrual. in a world of compute abundance (and therefore wider availability for training and inference) the demand continues to be centered on up the stack application and interface value accrual due to end user ownership.
when $/token effectively goes up, the consideration is about maximizing ROI (and therefore the highest margin allocation) through a balance between training and inference. there is no indication that there is and will be a slowdown in the demand for tokens - and therefore the highest ROI continues to be centered around deployment (and likely inference related workloads considering the performance of many of the frontier AI models widely available today).
the main positioning of how we think about the state of AI post midterms is primarily centered around the migration of value accrual. in a world of compute abundance (and therefore wider availability for training and inference) the demand continues to be centered on up the stack application and interface value accrual due to end user ownership.
when $/token effectively goes up, the consideration is about maximizing ROI (and therefore the highest margin allocation) through a balance between training and inference. there is no indication that there is and will be a slowdown in the demand for tokens - and therefore the highest ROI continues to be centered around deployment (and likely inference related workloads considering the performance of many of the frontier AI models widely available today).
0
0
3
208
there are parallels that we observe in software AI that will continue to emerge in robotics. some of those patterns will include:
1/ effective delegation of robot to robot collaboration. this primarily includes the ability to delegate tasks (sub-tasks) and goals
2/ human to robot collaboration. this includes both (1) delegation from humans and (2) wider collaboration on the ground.
3/ specialized models. post-training is accelerating (primarily due to @physical_int) and the release of API models @dreamlabs_ai. perhaps the key learning from software AI will carry over to rapid adoption of post-training specialized models across deployers.
4/ orchestrators responsible for delegating fleet management tasks. primarily draws on (1) but is truly a next "leg up," from the general fleet management software companies we see today.
where you land on this perspective is whether you believe there will be a (1) multi-hardware manufacturer and (2) multi-model future or a vertical integration from one core provider [a shining example of the latter being the figures of the world].
1/ effective delegation of robot to robot collaboration. this primarily includes the ability to delegate tasks (sub-tasks) and goals
2/ human to robot collaboration. this includes both (1) delegation from humans and (2) wider collaboration on the ground.
3/ specialized models. post-training is accelerating (primarily due to @physical_int) and the release of API models @dreamlabs_ai. perhaps the key learning from software AI will carry over to rapid adoption of post-training specialized models across deployers.
4/ orchestrators responsible for delegating fleet management tasks. primarily draws on (1) but is truly a next "leg up," from the general fleet management software companies we see today.
where you land on this perspective is whether you believe there will be a (1) multi-hardware manufacturer and (2) multi-model future or a vertical integration from one core provider [a shining example of the latter being the figures of the world].
1
0
5
334
significant announcement from apple today with the opening of developer integrations for Siri. core benefit is additional context that can be leveraged by Siri for personalized interactions.
in the current personal AI assistants environment, there are a few core components to s many argument: primarily associated with the need that individuals might have for > 1 assistant. is there a need for a productivity vs personal assistant? increasingly leaning to the former given the significant advancements in (1) model improvements and (2) the context switching required (including the introductions of new behaviors needed by users)
2/ the chat and voice interface: currently assistants are primarily chat-focused. increasingly clear the most natural interface that evokes "trust," and continued usage (as adoption) with AI is likely in the form of voice.
3/ ownership of the end user: in the race towards providing the one assistant to rule them all - how do you become increasingly defensible? apple is clearly angling the (1) hardware ownership. (2) muse's angle is primarily centered around context [personal context and shopping from FB, WhatsApp and their ecosystem of applications], and (3) familiar workflows which Instinct primarily leads with through the iMessage-first interaction
in the current personal AI assistants environment, there are a few core components to s many argument: primarily associated with the need that individuals might have for > 1 assistant. is there a need for a productivity vs personal assistant? increasingly leaning to the former given the significant advancements in (1) model improvements and (2) the context switching required (including the introductions of new behaviors needed by users)
2/ the chat and voice interface: currently assistants are primarily chat-focused. increasingly clear the most natural interface that evokes "trust," and continued usage (as adoption) with AI is likely in the form of voice.
3/ ownership of the end user: in the race towards providing the one assistant to rule them all - how do you become increasingly defensible? apple is clearly angling the (1) hardware ownership. (2) muse's angle is primarily centered around context [personal context and shopping from FB, WhatsApp and their ecosystem of applications], and (3) familiar workflows which Instinct primarily leads with through the iMessage-first interaction
0
0
6
453
we @decasonic are actively aiming to partner with founders in accelerating the deployment of physical AI at scale. we aim to invest in the app, interface, and services stack.
this includes:
1/ RL environments dedicated to (1) building task-specific expertise, (2) enabling human-robot collaboration on the ground, or (3) training and evaluating policies
2/ Inference companies aiming to tackle the on-device vs cloud orchestration question as physical AI systems potentially incorporate larger models and multiple specialized models (and orchestration of models becomes increasingly relevant)
3/ Machine coordination. we believe the future of physical AI includes specialized models and specialized hardware. this creates an opportunity for an interface that coordinates across hardware from different manufacturers and different types of models
this includes:
1/ RL environments dedicated to (1) building task-specific expertise, (2) enabling human-robot collaboration on the ground, or (3) training and evaluating policies
2/ Inference companies aiming to tackle the on-device vs cloud orchestration question as physical AI systems potentially incorporate larger models and multiple specialized models (and orchestration of models becomes increasingly relevant)
3/ Machine coordination. we believe the future of physical AI includes specialized models and specialized hardware. this creates an opportunity for an interface that coordinates across hardware from different manufacturers and different types of models
1
0
15
571
Abdul retweeted
We asked two isolated copies of the same AI to exchange a secret word with only a robot and a camera.
They succeeded by using air-writing, Morse code, flag semaphore and a 5-bit binary alphabet, without agreeing on a protocol beforehand.
They succeeded by using air-writing, Morse code, flag semaphore and a 5-bit binary alphabet, without agreeing on a protocol beforehand.
30
60
482
87,466
one of the main differentiating factors in the creation of consumer AI assistants today centers around user experiences
this is often focused on introducing new experiences that seamlessly translate into daily habits. two primary mechanisms for delivering these experiences found across both (1) vertical-domain assistants and (2) horizontal assistants are proactive AI agents and selective memory
1/ proactive AI is about the ability to provide a contextual, high-value experience that anticipates what matters to a user at a specific point in time
2/ selective memory is about recognizing what matters to a specific user and understanding how those preferences evolve over time. it is the foundation required to deliver deeply personalized experiences at scale
very excited for the future trajectory of the market. it is certainly a judgement heavy one!
this is often focused on introducing new experiences that seamlessly translate into daily habits. two primary mechanisms for delivering these experiences found across both (1) vertical-domain assistants and (2) horizontal assistants are proactive AI agents and selective memory
1/ proactive AI is about the ability to provide a contextual, high-value experience that anticipates what matters to a user at a specific point in time
2/ selective memory is about recognizing what matters to a specific user and understanding how those preferences evolve over time. it is the foundation required to deliver deeply personalized experiences at scale
very excited for the future trajectory of the market. it is certainly a judgement heavy one!
0
0
3
125
there was a rather interesting marketplace dynamic being played on openrouter where initial input token costs could be "undercut," by providers for certain models - with the same providers then overpricing for cached tokens.
@tarunchitra had a great piece on this issue: robvc.com/research/cachi…
@tarunchitra had a great piece on this issue: robvc.com/research/cachi…
1
0
1
206
the environment we are in increasingly demands AI-adopters to demonstrate their ROI from accelerated AI integration and spend.
one of the key unlocks from a company adoption standpoint is the aggressive storage and utilization of telemetry from AI systems:
(1) you first need to make sure you store your telemetry for the sake of observability
(2) you can sell logs in order to provide an "inference cashback," to data brokers if you are comfortable doing that (and you can always anonymize your own data before providing it to a vendor or broker)
(3) you can use your telemetry to build your own specialized models for both (1) recurring tasks and (2) judgement-heavy tasks with exceptions that need to be utilized and recorded
the key to benefitting from telemetry is ensuring an interconnected system-design approach. this requires the right (1) storage of information and (2) the effect identification of what matters to you from an ROI standpoint.
one of the key unlocks from a company adoption standpoint is the aggressive storage and utilization of telemetry from AI systems:
(1) you first need to make sure you store your telemetry for the sake of observability
(2) you can sell logs in order to provide an "inference cashback," to data brokers if you are comfortable doing that (and you can always anonymize your own data before providing it to a vendor or broker)
(3) you can use your telemetry to build your own specialized models for both (1) recurring tasks and (2) judgement-heavy tasks with exceptions that need to be utilized and recorded
the key to benefitting from telemetry is ensuring an interconnected system-design approach. this requires the right (1) storage of information and (2) the effect identification of what matters to you from an ROI standpoint.
2
0
5
339
Abdul retweeted
Learnings from today’s @decasonic AI Ship Day: Trust in AI grows when systems connect observed evidence, explicit assumptions, and human judgment to verified scenarios.
We learned from today’s demos that making that process visible and capturing the expertise that improves our thinking about today and tomorrow.
A few takeaways:
1. Forecasts and scenarios work together. A forecast assesses what is likely to happen. Scenarios explore what could happen under specific conditions. Connecting observations to milestones, a base case, and alternative scenarios helps us understand which assumptions drive the assessment and what evidence would change it.
2. Physical AI simulations become more useful with richer sensor data. We are active in deal flow across this area. Collecting and synchronizing data can create more signal-rich simulation environments. That richness helps capture physical interactions, operating constraints, and failure conditions. The opportunity is to build environments where AI can test scenarios against a more complete representation of reality, with real-world validation guiding each cycle.
3. Human observation and verification build trust. Chain of thought is part of this discussion: a model-generated explanation can help people investigate an answer, but it does not necessarily reveal the model’s actual internal reasoning or establish correctness. Trust requires traceable evidence, explicit assumptions, and verification. Experts identify meaningful signals, challenge missing context, and test whether the stated rationale holds up against observed outcomes.
4. Expert feedback can become valuable post-training data. As we train our open-source models, expert demonstrations and corrections can support supervised fine-tuning. Comparisons between outputs can support preference training, while reliably scored outcomes can provide reinforcement learning feedback. Capturing the context and rationale makes that feedback more useful. Improvement still needs to be demonstrated on cases held out from training.
5. Human expertise improves what reinforcement learning rewards. Experts help define success, distinguish meaningful progress from superficial performance, and identify when a system exploits a scoring rule without achieving the intended result. Capturing their interventions, corrections, and outcome assessments can improve reward models and evaluation criteria. Preserving who contributed the judgment and under what conditions helps us assess where that feedback applies. This gives RL a stronger signal for learning behavior that holds up in practice.
AI augmentation becomes more valuable when human expertise is captured, tested, and applied across workflows, esp those related to the future decisions we need to make. The opportunity is a learning loop where verified outcomes strengthen shared context and inform the next training cycle.
Always shipping. Always learning.
We learned from today’s demos that making that process visible and capturing the expertise that improves our thinking about today and tomorrow.
A few takeaways:
1. Forecasts and scenarios work together. A forecast assesses what is likely to happen. Scenarios explore what could happen under specific conditions. Connecting observations to milestones, a base case, and alternative scenarios helps us understand which assumptions drive the assessment and what evidence would change it.
2. Physical AI simulations become more useful with richer sensor data. We are active in deal flow across this area. Collecting and synchronizing data can create more signal-rich simulation environments. That richness helps capture physical interactions, operating constraints, and failure conditions. The opportunity is to build environments where AI can test scenarios against a more complete representation of reality, with real-world validation guiding each cycle.
3. Human observation and verification build trust. Chain of thought is part of this discussion: a model-generated explanation can help people investigate an answer, but it does not necessarily reveal the model’s actual internal reasoning or establish correctness. Trust requires traceable evidence, explicit assumptions, and verification. Experts identify meaningful signals, challenge missing context, and test whether the stated rationale holds up against observed outcomes.
4. Expert feedback can become valuable post-training data. As we train our open-source models, expert demonstrations and corrections can support supervised fine-tuning. Comparisons between outputs can support preference training, while reliably scored outcomes can provide reinforcement learning feedback. Capturing the context and rationale makes that feedback more useful. Improvement still needs to be demonstrated on cases held out from training.
5. Human expertise improves what reinforcement learning rewards. Experts help define success, distinguish meaningful progress from superficial performance, and identify when a system exploits a scoring rule without achieving the intended result. Capturing their interventions, corrections, and outcome assessments can improve reward models and evaluation criteria. Preserving who contributed the judgment and under what conditions helps us assess where that feedback applies. This gives RL a stronger signal for learning behavior that holds up in practice.
AI augmentation becomes more valuable when human expertise is captured, tested, and applied across workflows, esp those related to the future decisions we need to make. The opportunity is a learning loop where verified outcomes strengthen shared context and inform the next training cycle.
Always shipping. Always learning.
0
2
3
162
Abdul retweeted
Today, we're shipping real-time graphics.
Using Sol 6.1 Ultrafast we can generate custom overlays, diagrams, and even interactive 3D scenes right on a user's screen, no chat.
It's a new visual interface for AI.
Here's me learning about LLMs + hardware.
Demo:
Using Sol 6.1 Ultrafast we can generate custom overlays, diagrams, and even interactive 3D scenes right on a user's screen, no chat.
It's a new visual interface for AI.
Here's me learning about LLMs + hardware.
Demo:
704
841
11,700
1,256,364
benchmarks from @robocurve are being used to measure real world economic tasks.
this will serve as an inflection point from the movement of running evals for generalized robotics tasks into tasks with economic consequences, which will (1) continue to accelerate as more robots ll allow for the further deployment of robots at scale.
this will serve as an inflection point from the movement of running evals for generalized robotics tasks into tasks with economic consequences, which will (1) continue to accelerate as more robots ll allow for the further deployment of robots at scale.
DataCenterBench Tier 1 is now on @FireworksAI_HQ’s Specialized Intelligence Index.
We evaluate robots on data-center maintenance tasks, like connecting cables and operating controls, to track progress toward automating the physical work behind compute.
Explore the results: fireworks.ai/specialized-in…
0
0
3
196
apple announced the release of macOS 27 'Golden Gate.' there's a clear interface expansion here and a potential hint on the future trajectory use-cases offered by the intelligence feature:
(1) Siri as the on-device AI assistant, with context being drawn in from across both your (multi-device)
(2) there seems to be a potential use-case focused on agentic commerce in the near-future from apple. one of the main use-cases highlighted (for relevant AI workflows) refers to: "product restocks or price drops," which points towards a potential consideration for the expansion capabilities of Apple's Intelligence ecosystem (beyond the utilization of Siri)
the main question surrounding assistants continues to be primarily centered around the point of value accrual:
1/ does the owner of the hardware own the value? re Apple
2/ does the owner of personal context? gemini
3/ does the owner of social-focused context? meta
4/ or potentially new companies aiming to win through differentiated features (1) multiplayer, (2) group-chats, and (3) pro-active AI assistants
(1) Siri as the on-device AI assistant, with context being drawn in from across both your (multi-device)
(2) there seems to be a potential use-case focused on agentic commerce in the near-future from apple. one of the main use-cases highlighted (for relevant AI workflows) refers to: "product restocks or price drops," which points towards a potential consideration for the expansion capabilities of Apple's Intelligence ecosystem (beyond the utilization of Siri)
the main question surrounding assistants continues to be primarily centered around the point of value accrual:
1/ does the owner of the hardware own the value? re Apple
2/ does the owner of personal context? gemini
3/ does the owner of social-focused context? meta
4/ or potentially new companies aiming to win through differentiated features (1) multiplayer, (2) group-chats, and (3) pro-active AI assistants
0
0
0
167
this is quite the interesting read from the @GoogleDeepMind team, and it speaks to areas of adoption for the intersection of AI and crypto.
basis of the article is primarily around trust in evaluation through benchmarks. as an eval company, you can either (1) handoff your ume trust in the model developer or (2) receive model weights from the model developer.
there's friction in both approaches - and Google's 'Confidential Computing Portfolio,' aims to cryptographically verify both the external evaluation data (i.e the prompts) and the model weights remain owned by the respective parties.
basis of the article is primarily around trust in evaluation through benchmarks. as an eval company, you can either (1) handoff your ume trust in the model developer or (2) receive model weights from the model developer.
there's friction in both approaches - and Google's 'Confidential Computing Portfolio,' aims to cryptographically verify both the external evaluation data (i.e the prompts) and the model weights remain owned by the respective parties.
0
1
6
252
in speaking with companies actively (1) adopting AI or (2) seeking to work with AI-native portfolio companies through our due diligence - there's an over-emphasis on existing behaviors and state.
there's often a bias towards introducing AI for (1) automation and (2) for replacement of an existing process. there's no indication or potential consideration that there often times - has to be a consideration of a clean slate in order to account for the full range of possibilities introduced by AI.
imo - this is the "chaos," that @Replit alludes to. specifically with companies like Replit, they provide the ability to introduce personalized, sometimes one-off software at scale. chaos from the discovery, maintenance, and improvement associated with the one-off software can be painful to an enterprise. a focus on autonomous enterprises could "de-risk," the chaos - as AI is great at transforming the unstructured to structured.
there's often a bias towards introducing AI for (1) automation and (2) for replacement of an existing process. there's no indication or potential consideration that there often times - has to be a consideration of a clean slate in order to account for the full range of possibilities introduced by AI.
imo - this is the "chaos," that @Replit alludes to. specifically with companies like Replit, they provide the ability to introduce personalized, sometimes one-off software at scale. chaos from the discovery, maintenance, and improvement associated with the one-off software can be painful to an enterprise. a focus on autonomous enterprises could "de-risk," the chaos - as AI is great at transforming the unstructured to structured.
0
0
2
253
very interesting to see how @SpaceX is consolidating their AI interface play:
(1) for now it's primarily a combination of Grok Bot + Cursor.
(2) what's really interesting is how actively they are pushing Grok Bot into everything. available in your DMs, is connected to X Money account (for agentic finance use-cases), and is active in the standalone desktop application
(3) if there's persistence across those interfaces (memory and context primarily); Grok could potentially cross the chasm of offering both personal and professional targeted use-cases for consumers (which I think provides an additional layer of defensibility considering the recent prevalence of those applications)
(1) for now it's primarily a combination of Grok Bot + Cursor.
(2) what's really interesting is how actively they are pushing Grok Bot into everything. available in your DMs, is connected to X Money account (for agentic finance use-cases), and is active in the standalone desktop application
(3) if there's persistence across those interfaces (memory and context primarily); Grok could potentially cross the chasm of offering both personal and professional targeted use-cases for consumers (which I think provides an additional layer of defensibility considering the recent prevalence of those applications)
4
0
10
737
Abdul retweeted
LLMs have a score doubling time of 53.9 days on a robotics benchmark between July and October 2026, while speed increased by 4.4x and cost fell by 6.2x.
More results in our new paper: The Bitter Lesson Comes to Robotics 🧵
More results in our new paper: The Bitter Lesson Comes to Robotics 🧵
26
53
454
82,277
increasingly, there's a clear-trend of companies emerging with the dedicated goal of selling real world deployment data (logs and traces - "data") to frontier labs. the business model is primarily predicated on determining the (1) value of benchmark improvements relative to the (2) cost of acquiring the data.
there's a very interesting question about the number of middle "layers," needed to provide the data back to the labs. there's a potential opportunity in the future where labs could offer "inference cashback," beyond the "promotional period," often seen on @OpenRouter and other inference marketplaces that could last and be durable. this cashback would be offered to businesses and users directly, with the labs guaranteeing the "anonymization," of user data.
will see if this emerges in the near-future, but there's clearly an opportunity for "capital markets," in the data layer to play a role.
there's a very interesting question about the number of middle "layers," needed to provide the data back to the labs. there's a potential opportunity in the future where labs could offer "inference cashback," beyond the "promotional period," often seen on @OpenRouter and other inference marketplaces that could last and be durable. this cashback would be offered to businesses and users directly, with the labs guaranteeing the "anonymization," of user data.
will see if this emerges in the near-future, but there's clearly an opportunity for "capital markets," in the data layer to play a role.
3
1
28
1,677