The Second History of Intelligence

Our Mission

The mission of SADAIR is to make intelligence a lasting human capability.

AI’s first history expands what models can do. Its second is taking shape wherever people and organizations put that intelligence to work. We believe they should be able to build it into their own environments, shape it with their knowledge, and develop uses that no central model builder could have planned. The models will keep improving. The tools, practices, and understanding formed through their use should give people a lasting capacity to create.

Intelligence acquires direction from purposes that are not its own. People choose those purposes, weigh the tradeoffs, and remain answerable for the consequences. Civilization preserves what people manage to record, while much of the judgment that guides action leaves no durable trace. A manual can preserve a procedure; practice teaches when that procedure should bend. Experience leaves documents and changed people. Skill, taste, relationships, and practical judgment live in those people and continue to change with them. General models offer a common inheritance. Through use, intelligence develops a personal history.

The knowledge that gives intelligence direction is dispersed across people and organizations. Much of it emerges through work, as people encounter a problem, reconsider a goal, or recognize that an established method no longer fits. General models provide a growing common foundation. People need the means to bring their own judgment into those systems and adapt them to changing purposes. The ability to shape intelligence should be as widely distributed as the knowledge and purposes it must serve.

That distribution can change how progress is made. People working in different environments can explore different possibilities, test them against actual needs, and learn from the results. A need visible only in one setting can become the beginning of a useful tool or a new practice. Those discoveries can travel between settings and help improve the common foundation. We want a society in which more people can contribute to the development of intelligence, and in which technical progress expands humanity’s ability to create and pursue its own purposes.

SADAIR is building a Personal Intelligence Layer to give that participation a technical foundation. It has a local center and a frontier reach. On-device language and voice models carry the continuous work of interaction and everyday inference. People and organizations can bring their own knowledge into that work and direct how the system acts within their environments. Frontier models remain available for deeper reasoning, specialized knowledge, and compute beyond the device. People and organizations retain control of the context and corrections formed through use, so they can build on earlier work as the shared models improve.

We are building this foundation for people to use directly and for developers to bring into their own software, devices, and machines. In a workplace, personal judgment and organizational knowledge can shape the same task, with distinct permissions for private history and shared work context. The same foundation can support a personal assistant, an application, a pair of glasses, a robot, or an instrument, with perception and action bounded by each product’s domain-specific controls.

A common foundation should give people room to develop different forms of intelligence around the lives they lead and the work they choose. Four commitments organize our work.

• We distribute intelligence to the edge. Compact language and voice models turn device compute into reusable intelligence near the places where perception, context, and action originate. We design for a device-complete system, in which every capability, from inference to voice to learning, can run on the device and the cloud extends it. Local execution changes both availability and economics. Interactions completed on device have zero inference cost, while frontier infrastructure remains available for work that requires greater scale.
• We make on-device intelligence conversational. Conversation carries meaning through timing, tone, hesitation, overlap, silence, and revision. We build speech generation and real-time interaction control together so people can interrupt, correct, and redirect the exchange while it unfolds.
• We connect personal and frontier intelligence. The local system handles the continuous path and reaches outward when a task requires stronger reasoning, current knowledge, specialized tools, or elastic compute. Permissioned routing lets frontier capability evolve independently while personal state, policy, and continuity remain anchored on the user’s side.
• We measure progress in human capability. Model benchmarks remain essential. Tasks with clear objectives and reliable safeguards can often be automated. For work shaped by judgment, the complete system must also be judged by what people and AI can accomplish together, how well people can correct and direct the work, and what knowledge, skill, and agency remain with them over time.

The Personal Intelligence Layer is our contribution to that second history.

The Second History

The history of a general-purpose technology can be understood in two movements. The first establishes a new capability and raises its frontier. The second distributes that capability, changes the economics of using it, and gives societies time to rebuild products, organizations, and everyday practice around it. Carlota Perez describes the shift from the installation of a technological revolution to its deployment, when “the fabric of the whole economy is rewoven and reshaped.” We use second history for this longer work. The first history makes a new power possible. The second shapes how many people can exercise it, how many purposes it can serve, and what kind of civilization can grow around it.1

Electricity’s first history made large-scale generation and transmission practical. Power stations and utility grids separated production from the point of use and carried current into streets, factories, and homes. Yet many early factories used a large electric motor to drive the existing shaft-and-belt system, preserving the old floor plan and organization of work. The source of power had changed. The architecture of work had not.2

The deeper transformation came through the unit drive. Smaller motors powered individual machines, so equipment could be arranged around the sequence of production. One machine could start, stop, or move without forcing the rest of the factory to follow. Engineers and managers rebuilt machinery, layouts, workflows, and organizations around choices that steam had never allowed. The grid distributed a common capability. Thousands of factories discovered what that capability meant through use.

Computing followed the pattern through a different architecture. Mainframes established computation as an institutional resource, and time-sharing widened access while the machine, its administration, and most of its state remained elsewhere. Personal computers placed processing, memory, software, and a persistent workspace under an individual’s control. Alan Kay and Adele Goldberg imagined a responsive medium that could carry its owner’s information and become many different tools. Mark Weiser later described mature computing as something woven into ordinary life.3

Central computing continued to grow. Networks and data centers expanded alongside personal devices. The personal computer supplied a durable point of control from which people could create software, connect to larger systems, and invent uses that no central designer could have enumerated in advance. Central scale and personal capacity became complementary layers. Their interaction became the foundation for new products, industries, media, and forms of work.

AI now stands at a comparable threshold. Its first history built the mathematics, hardware, data, training methods, and institutions behind systems that can perceive, reason, generate, and act across an expanding range of tasks. Training frontier models still requires concentrated talent, capital, data, and compute, and that work should continue. Advancing the frontier establishes what is possible. Its wider consequences depend on whether people and organizations can use that capability repeatedly, afford its use, govern its role in their work, and build their own practices around it. Cloud services have already widened access to intelligence at extraordinary speed. Access determines who can call on intelligence. Authority determines who can preserve the state formed through use, govern its permissions, and carry that continuity forward. We believe AI’s next distribution will place a growing share of that capacity near the user, where suitable models can run repeatedly, act within a local environment, and call outward when stronger intelligence is needed.

The architecture emerging from this transition is local-first and cloud-extended. Compact models and local inference frameworks supply low-latency perception and reusable inference near the places where people speak, work, and act. User-side state preserves continuity and governs private context. Frontier infrastructure supplies broad knowledge, deep reasoning, specialized capability, and elastic compute. For inference completed entirely on device, the inference cost is zero. Once a model and compatible inference framework are installed, repeated use draws on reusable device capacity instead of a metered service. The economic unit changes from a cloud request purchased one call at a time into intelligence that can be owned and used again. High-frequency software and background agents can invoke that capacity throughout the day without a token bill rising with every event.4

Paired with a streaming interaction architecture, local capacity also changes the unit of interaction. Within an active, permissioned session, the system moves beyond a fixed prompt-and-response cycle into an event-driven, interruptible live loop. Full duplex allows it to keep listening while it speaks. Partial speech, interruptions, and returning background results can change what it does next. Turn-taking becomes a continuous prediction, allowing the system to acknowledge, yield, and recover as the exchange unfolds. By always-on, we mean continuous availability during that user-controlled session, with visible controls for pause, mute, retention, and remote routing.

That reusable capacity and live interaction loop widen the application frontier. Developers can build intelligence into software and machines used throughout the day. People and organizations can put those systems to work, discover where they fall short, and help their builders revise them. As the systems improve, people can reorganize the work around capabilities that were previously too costly or impractical to use continuously. Each domain still imposes its own hardware and safety requirements. David Edgerton’s use-centered history brings the places where technology is used into view.5 When history records only where models are trained, AI belongs to a few centers. When it records where intelligence acquires a use, nearly every place can enter the story.

Here AI departs from electricity and computing. Electricity can deliver the same current to many machines, and computing can expose the same primitives to many programs. Intelligence takes direction from the knowledge, purposes, relationships, and constraints it encounters in use. A common model may arrive with broad capabilities, while a person or organization supplies distinctions that no general training corpus could contain in advance. Some of that knowledge is private, some is difficult to state, and some changes as people act, observe the result, and revise their judgment. As intelligence enters different lives, the distribution of intelligence can become the diversification of intelligence.

That diversity has an economic role. Eric von Hippel’s research shows that users often create important innovations because detailed knowledge of a need and its context is difficult to transfer to a producer who does not share that use context. User innovation then becomes a complement and an input to manufacturer innovation.6 His mechanism suggests what distributed AI capacity can change. When people and organizations can adapt intelligence close to use, they can discover tasks, tools, workflows, and products that a central model developer could not specify beforehand. The point of use becomes a point of invention.

General-purpose technologies also improve through feedback between the common capability and the applications built around it. Timothy Bresnahan and Manuel Trajtenberg called this relationship innovational complementarities. Improvements to the underlying technology open new applications. The growth of those applications increases the value of improving the underlying technology again.7 AI’s second history can therefore feed its first when discoveries circulate through consent, common standards, and institutions capable of coordinating the two. Frontier labs raise the common ceiling. Distributed environments enlarge the search space of needs, failures, methods, and evaluations from which the next generation of intelligence can learn.

Distributed intelligence would let more people and institutions conduct more experiments at once across enterprise, education, care, science, art, and collective organization. The result could be a wider field of productive activity and a society better able to form knowledge, compare judgments, and explore possible futures. Claude Lévi-Strauss wrote that “all cultural progress depends on a coalition of cultures.”8 Difference becomes generative when distinct forms of knowledge can meet, challenge one another, and combine. A plural ecology of intelligence needs local centers that preserve what is particular alongside common networks through which discoveries can travel.

Access to intelligence and authority over its use are different properties. A model can run on a personal device while its state, update path, and permissions remain under a distant provider’s authority. Meaningful distribution requires a durable center on the user’s side, with control over continuity, disclosure, adaptation, portability, and action. Local-first software established this authority pattern for data by treating the user’s local copy as primary while network services support synchronization and collaboration.9 Personal intelligence extends the pattern across models, memory, interaction, and action.

Distribution creates a coordination problem. Without durable continuity, a person’s history would fragment across devices and services. Without frontier reach, personal systems would be cut off from the strongest shared intelligence. The Personal Intelligence Layer connects these scales. It gives each person or organization a stable center from which local and frontier systems can act together, while preserving the differences through which intelligence continues to develop.

The first history raised the frontier of intelligence. The second will determine how widely the power to use, shape, and govern intelligence is distributed. When the two reinforce one another, a common technical inheritance can widen the range of futures humanity has the capacity to create.

The Personal Intelligence Layer

The system has two execution paths. The live path stays on the user’s device and can remain continuously active under the user’s control. A Personal Language Model and the voice models execute through our on-device Inference Framework, while a real-time Voice Controller independently governs acoustic timing, playback authority, and interruption. Together they process authorized signals from speech, timing, application events, and device context as they arrive. They maintain a provisional view of the current interaction and read confirmed information from the Memory System. On-device interactions have zero inference cost and continue when the network is unavailable.

When the Personal Language Model identifies a committed need for deeper reasoning, current knowledge, or a tool unavailable locally, it creates a provider-blind remote delegation request. The request specifies the capability needed and leaves provider selection to the Router. The Context Compiler prepares the permitted candidate views, and the Smart Router selects one remote contract for the background path. The live path continues while that task runs. Returned results carry a source and task revision so the system can discard stale work before it enters the conversation.

Live interaction and the Memory System follow separate lifecycles. Partial transcripts, timing signals, and inferred intent remain provisional. Confirmed goals, corrections, permissions, and unfinished work can persist across sessions. A remote result cannot update the Memory System, speak, or execute an action until the local policy layer accepts it.

Personal Language Model

We built SADAIR’s Personal Language Models for developers bringing on-device reasoning and tool use into their applications. PLM 1.8B prioritizes response speed and a small memory footprint. PLM 3B offers stronger capability in a compact model, while PLM 8B-A3B provides the family’s highest capability for applications with a larger memory budget. All three are designed for repeated, batch-one inference. Our aim is to make useful language intelligence a resource software can draw on throughout a task, without sending every interaction to the cloud.

Each model has Base and Instruct checkpoints, a shared tokenizer, and a trained 128K context window. A common output protocol lets developers change model size without rebuilding the surrounding application. Within the Personal Intelligence Layer, PLM interprets requests, produces structured outputs, and proposes remote assistance when a task needs greater capability. Acoustic timing, route selection, and durable user state remain outside the checkpoint.

Model architecture

The architecture combines double-gated causal ShortConv for local sequence mixing with grouped-query attention (GQA) for access across the wider context. ShortConv keeps a fixed-size convolutional state. Only the attention layers accumulate a growing KV cache, so 25–27% of layers incur that sequence-length-dependent memory cost. We selected the layer layouts through ablations run on the target devices, weighing model quality against response time, memory, and sustained execution. The family figure shows how this balance changes with model size.

PLM 1.8B and PLM 3B use dense SwiGLU feed-forward layers. PLM 8B-A3B expands model capacity through sparsity, with dense FFNs in the first two decoder layers and mixture-of-experts FFNs in the remaining twenty-two. A normalized-sigmoid router sends each token to its top-4 of 16 experts in each sparse layer, with no token dropping. This gives the model 8.47B total parameters with 2.65B active per token. Sparse activation reduces the computation required at each step; all expert weights still occupy memory. The architecture below makes both the sequence-mixing and expert paths explicit.

Training

Our training strategy invests compute up front to improve the capability of models that remain compact at deployment. Each Base checkpoint processed 29.8T non-padding token-passes drawn from a shared corpus containing 10.2T tokens after deduplication, spanning reference text, long-form documents, code, mathematics, structured data, and verified synthetic material. High-quality sources are resampled across pretraining, long-context mid-training, and distillation. Data preparation includes rights and quality review, PII and credential filtering, exact and semantic deduplication, and benchmark decontamination. This inference-oriented overtraining spreads a larger training investment across repeated uses of a small deployed model.

The family shares a 128,256-token byte-level BPE vocabulary, including 256 reserved protocol tokens. The vocabulary improves compression across prose, code, mathematics, and structured inputs; reversible byte fallback preserves arbitrary UTF-8 without implicit Unicode normalization. The initial release will support English and Spanish, with dedicated post-training and evaluation in both languages.

Long-context mid-training progresses through 32K, 64K, and 128K sequence stages using complete documents, repositories, multi-document tasks, and workflow histories. Short-context examples remain in the curriculum throughout. RoPE with θ = 1,000,000 and the 131,072-position limit are fixed at initialization, so the positional scheme stays consistent as training sequences lengthen. The trained 128K context window gives applications room for extended working context, while the hybrid architecture limits the number of layers that must cache it.

Distillation from teachers specialized in general language, code, and mathematics concentrates broader capabilities into the compact models. Specialist examples pass deterministic verification before entering training. We allocate more distillation to the smaller checkpoints, where limited capacity makes teacher supervision especially valuable. Distillation targets general language modeling; instruction behavior and tool protocols are introduced during post-training.

Post-training teaches the models to complete a task and to recognize when the available evidence or capability is insufficient. We combine rejection-sampled supervised fine-tuning, preference optimization, and reinforcement learning with outcome-verifiable and calibrated rubric-based rewards. Executable verifiers take priority, with calibrated judges for open-ended tasks and human review of disputed or high-risk examples. A capability-boundary curriculum uses 10 million shared counterfactual task forks and 100,000 human-written or adjudicated seed episodes. The task stays fixed while evidence, ambiguity, freshness, tools, permissions, and difficulty change. The appropriate response may be to answer, clarify, use a local tool, request a provider-blind remote capability, or abstain. Training penalizes unsupported answers and unnecessary escalation alongside missed escalation and blanket refusal. A final safety-consolidation stage completes each Instruct checkpoint.

Model capabilities

We evaluate whether PLM can follow explicit instructions, reason through mathematical problems, form valid tool calls, and carry a task through to completion. IFEval, MATH-500, BFCL v4, and τ²-Bench Telecom test these capabilities separately. Together they cover several demands of an application that relies on a local language model for both conversation and action.

We measured the four capability benchmarks below in English using BF16 Instruct checkpoints under a common evaluation protocol. Models had access only to the tools permitted by each benchmark. The Smart Router and remote assistance were disabled. The comparison therefore measures each model’s own capability within the benchmark environment.

PLM 8B-A3B scores 90% on IFEval, the highest in this comparison, and 86% on MATH-500. It combines strong instruction following with competitive reasoning and tool use for applications that need to carry a request through several steps. All four scores improve as the family scales from 1.8B to 3B to 8B-A3B. The device results below show the response time and memory footprint alongside that progression, so developers can choose a model for the work their application needs to do.

The surrounding conversation places further demands on the model. A correction should change the response without losing the task, and an expression of frustration should inform how the model responds without becoming an assumption about the person. The example below shows how we assess these choices at the level of an individual response.

On-device inference

For mobile applications, the smaller PLMs make room for language intelligence alongside the rest of the product. On iPhone 17 Pro, PLM 1.8B has a time to first token (TTFT) of 0.67 seconds and decodes at 65.0 tokens per second, with peak memory of 1,598 MB in the separate 4K-context test. PLM 3B decodes at 43.0 tokens per second while using 35% less peak memory than Qwen3.5-4B. Developers can trade some of the smaller model’s speed for the stronger capabilities shown above without taking on the largest model’s memory footprint.

Our mixed-precision W4 and W8 deployments use low-bit weights for large projection matrices and retain higher precision for sensitive operations, including routing, normalization, and convolution. Recovery tests determine the precision of each part. The Inference Framework pairs packed matrix operations and grouped expert dispatch with fused gated convolution and GQA state management. Cancellable generation and controlled prefill let the surrounding system interrupt or schedule work as a task changes.

The mobile comparison uses Q4_K_M-class mixed precision and a common llama.cpp backend within each device, using Metal on iPhone and OpenCL on Galaxy. Each timed request processes 1,024 input tokens and generates 256 output tokens at batch one, without prefix-cache reuse. TTFT excludes model loading and initial compilation. These are single-model text measurements, separate from concurrent voice execution and our production Inference Framework; the capability benchmarks above use BF16 checkpoints. The separate PLM 8B-A3B comparison follows the measurement conditions stated in the figure.

On Galaxy S26 Ultra, PLM 1.8B also has the lowest TTFT, fastest decode, and smallest peak memory in the mobile comparison. PLM 3B decodes faster and uses less memory than Qwen3.5-4B, though its TTFT is 3.78 seconds, compared with 2.93 seconds for Qwen. The two sizes offer different balances of capability, TTFT, and decode speed.

PLM 8B-A3B serves applications that can allocate more memory to language capability. In its separate comparison with Qwen3.5-4B and Gemma 4 E4B, it has the fastest decode and shortest completion time on all three devices, with peak memory of roughly 5.7 GB at 4K context. All expert weights remain resident even though only a subset executes for each token. The family gives developers a common foundation across these deployment choices, from compact phone applications to more demanding local workloads.

Future Work

Our next phase of PLM research focuses on lowering response latency, reducing the memory and energy required for local inference, and enabling personal intelligence to learn from experience. Greater efficiency can leave more of the device available for voice, perception, and the application itself.

• Lower-bit inference. We are exploring a separately trained, native ternary PLM 1.8B using W1.58A8 computation and dedicated inference kernels. We study training methods and execution paths together, with the aim of reducing memory use and energy consumption while improving real-device speed. We evaluate instruction following, tool use, contextual reasoning, and safety against our qualified W4 deployments. Our longer-term goal is to make 10B-class and eventually 20B-plus models practical on capable smartphones as low-bit methods, inference kernels, and mobile hardware advance. This would expand the capability available on personal devices while preserving the quality required for everyday work.

• Learning from personal experience. Our experimental program examines how memories, corrections, and task outcomes can become training data with the user’s explicit approval. The system selects and constructs this data so people can teach it through ordinary use. We study Self-Distillation Policy Optimization to learn from interaction feedback, Self-Distillation Fine-Tuning to consolidate demonstrated behavior, and Doc-to-LoRA-style methods to generate lightweight adapters from context. Across these approaches, we examine what should remain in editable memory and what should change the model’s weights. We test how to distinguish lasting preferences and transferable skills from temporary instructions, so experience can improve future behavior across tasks and sessions.

• A system-level path to continual learning. In the medium term, our approach combines memory management, automated data construction, cloud post-training on examples the user explicitly approves for off-device training, and versioned model deployment. We are exploring how Thinking Machines Lab’s Tinker can support this work as it develops. Our experiments track adaptation across successive updates, measuring gains on personal tasks alongside retention of general capabilities and safety. We evaluate candidate updates before deployment and preserve a rollback path to the previous model. The system should carry the burden of data preparation, training, and validation, while people decide which experiences it may learn from and approve changes to their model.

Over the longer term, we aim to develop models and learning methods that acquire new skills from fewer interactions, retain established capabilities, and adapt as people’s needs change. A common model foundation should support many paths of development, shaped by different people, practices, and cultures. We see continual learning as a path toward intelligence that develops through use. It can broaden whose knowledge shapes AI and preserve the differences from which new ideas emerge. Human experience should become a lasting source of progress in intelligence.

KRORAINA 2

In a spoken exchange, people can redirect the work before an answer is finished. A pause may leave a thought open; a brief acknowledgment can encourage the speaker to continue. KRORAINA 2 brings contextual speech generation and real-time interaction control to on-device applications. We design it so people can shape an exchange as it unfolds, with a voice system that can keep listening while it speaks.

A technician consulting instructions through glasses paired with a phone needs to ask a follow-up question or interrupt an explanation without putting down their tools. For robotics developers, the corresponding opportunity is a conversational interface for giving and revising instructions, with physical actions governed by the robot’s own control and safety systems.

Edge is the phone-first profile, designed for sustained real-time speech. Quality allocates more capacity to acoustic detail on Mac and PC. Both use system-level full duplex, with the microphone and renderer active together. The Personal Language Model commits the main response, KRORAINA 2 generates its spoken delivery, and the Voice Controller governs playback. This separation lets the system acknowledge, yield, and recover during an exchange. The local voice path runs without metered cloud inference.

The profiles share a context format and streaming interface and support en-US and es-US. English–Spanish code-switching is evaluated separately. The generator counts above exclude the profile-specific waveform decoders, input audio encoder, and control models. The listening examples below let readers hear how the system waits, acknowledges, yields, and repairs an interrupted response.

Contextual speech generation

How a sentence should sound depends on the exchange around it. KRORAINA 2 conditions each segment on recent text and audio, an authorized speaker-identity anchor, its own preceding delivery, and the committed words to speak. Pronunciation and delivery controls share the 2,048-position multimodal context with those inputs and the generation budget. Delivery can therefore respond to changing context while the model’s weights remain fixed.

The generator divides its work across two timescales. A Temporal Transformer predicts the first codebook of each 80-millisecond frame, carrying linguistic and prosodic structure across frames. A smaller Depth Transformer completes the acoustic detail within each frame, predicting the remaining seven codebooks for Edge and fifteen for Quality. Both use the frozen codebooks of the Mimi encoder and RVQ quantizer for 24 kHz audio.

For efficient streaming reconstruction, we replace Mimi’s decoder with independently trained iSTFT waveform decoders of 16M parameters for Edge and 29M for Quality. Each predicts spectral magnitude and phase from the completed codes, then reconstructs audio through an inverse short-time Fourier transform. Keeping the input encoder, quantizer, and codebooks fixed lets us optimize waveform synthesis while preserving the generator’s audio representation and the control models’ inputs.

We trained the generator family on approximately 933,000 hours of speech after filtering and split assignment, with dedicated training views for high-quality audio, intact sessions, and two-person conversation. Our own recording program contributed 1,210 session hours from 112 professional actors in 70 stable dyads, varying identity, emotion, and delivery independently. A separate synthetic pipeline retained 3,600 hours before splitting to cover rare counterfactual, safety, echo, and code-switch cases. Real conversational recordings supply the evidence for natural human timing.

We trained Quality from random initialization through general pretraining, continued pretraining on high-quality audio, contextual and dyadic mid-training, actor and timing refinement, and preference optimization. Context training pairs the same target waveform with matched, absent, shuffled, stale, and same-actor-wrong-dialogue histories. A ranking objective favors the matched history while holding text and speaker identity fixed. Edge was distilled from the frozen Quality teacher before independent post-training. Mobile optimization retains these Temporal backbones and adapts the Depth and waveform networks separately.

Both profiles use Slim Depth, a distilled Depth Transformer with a narrower feed-forward network. Figure 10 follows the generation path from dialogue context to waveform; switching profiles reveals how we allocate capacity within the same streaming interface.

On-device speech inference

On iPhone 17 Pro, our mobile optimizations reduce Edge’s median real-time factor (RTF) by 31%. We target the work repeated within every speech frame. Slim Depth reduces feed-forward width from 8,192 to 3,072 in Edge and 4,096 in Quality, preserving layer counts, attention dimensions, and codebook order. Distillation transfers the full-width model’s code predictions into the compact network. The waveform decoders learn from source audio through reconstruction, spectral, and adversarial losses. Streaming recovery training covers chunk joins, pauses, and utterance endings along the same stateful path used for playback.

We allocate precision by module, using W4A16 for the Temporal backbone and W8A16 for Depth, with sensitive output heads in FP16. Waveform-decoder matrices use W8 while spectral outputs and overlap-add reconstruction retain higher precision. Calibration and recovery training cover both languages, with particular attention to expressive speech and long utterances where delivery is most sensitive to quantization.

Metal and Vulkan execution preserve convolution, attention, and synthesis state across chunks. Preallocated Depth buffers and fewer CPU–GPU synchronizations reduce per-frame overhead, while the cached attention state (KV cache) avoids recomputing earlier context. We measure these changes together under the same streaming workload, then test whether the gain survives a sustained session. That distinction determines which profile belongs on a phone.

Real-time interaction control

Conversation should help people stay with their work, without making the interface another task to manage. A person should be able to pause without losing their turn and interrupt a response when the work changes. KRORAINA 2 combines predictions about the conversation with deterministic playback control, so it can keep listening while deciding whether to wait, acknowledge, or yield.

KRORAINA 2 uses fewer than 17 million control-model parameters beyond the shared codec encoder. KR-VAP (3.74M parameters) predicts activity across the next two seconds from both sides of the exchange. KR-EOT (8.19M parameters) is an event-triggered, audio-native end-of-turn model. It estimates whether the person is yielding the floor or continuing, with an explicit UNCERTAIN outcome, without waiting for final ASR punctuation. These predictions inform playback decisions but cannot authorize them.

When speech begins during playback, an Acoustic Guard can request reversible ducking while an interruption classifier, KR-Interrupt, gathers evidence. The classifier distinguishes an attempt to take the floor from cooperative overlap, a listener backchannel, laughter, third-party speech, echo, and noise. A listener-response policy, KR-BC, can remain silent or propose a brief continuer, neutral acknowledgment, affective reaction, or repair bid. The generator realizes an accepted response in context; training penalizes responses that make unwarranted claims of understanding or agreement.

We trained KR-VAP on 38,400 hours of synchronized two-person conversation. KR-EOT uses 19.2 million training cutpoints drawn from a 12,000-hour registry that includes 600,000 triple-annotated consensus examples. The listener-response policy learns from 1.2 million training opportunities, with sparse positive responses and examples that teach silence or reject inappropriate insertions. Training, development, and sealed evaluation sets remain separate, with speakers, dyads, sessions, scripts, rooms, and derived windows confined to the same split. These datasets teach different judgments, from predicting speech activity to recognizing when silence is the appropriate response.

Only the deterministic Voice Controller can authorize playback or change the playback state. It keeps listening when KR-VAP and KR-EOT disagree. Acoustic echo cancellation (AEC) and target-speaker verification remain active during playback. Confirmed floor-taking cancels speech that has not yet been played; a backchannel from the user can permit playback to resume. Repair follows the current task revision and the hardware-timestamped estimate of what was played.

Generation advances in 80-millisecond frames, while the renderer commits audio in 10-millisecond blocks. Ducking and playback revocation take effect at that finer boundary. The system can therefore lower its voice while gathering evidence for cancellation, without waiting for another generated frame. Cancellation discards the buffered tail so an interrupted response cannot continue playing from the queue.

Figure 11 shows how acoustic evidence reaches the Controller. Figure 12 follows those decisions through a conversation, from an unfinished pause to an interruption and repair. The distinction between predicted events and authorized playback remains visible throughout.

Voice evaluation

We evaluate how KRORAINA 2 sounds, when it speaks, and whether it can sustain a live exchange on device. Listening studies assess delivery, controller ablations isolate timing decisions, and device measurements test the cost of running the voice stack continuously.

For speech quality, a fixed ASR scorer and normalization procedure measure word error rate (WER), while a fixed speaker encoder measures speaker similarity. English, Spanish, and code-switching are scored separately. Blind listening compares each profile with human recordings under matched text, voice identity, and presentation conditions. The charts report 95% confidence intervals and judgment counts; preference percentages exclude ties.

WER is close between profiles in each language condition, while Quality has higher measured speaker similarity. In blind listening, Quality receives 47.7% of preferences against human recordings, with a 95% interval of 44.2–51.2%, so the study shows no clear preference between Quality and the human recordings. Edge receives 41.9%, with listeners preferring the human recordings. The device measurements below show the compute budget needed to sustain each profile in conversation, allowing developers to weigh voice fidelity alongside response time and memory.

For interaction timing, we compare the VAD/silence baseline, KR-VAP, KR-VAP with KR-EOT, and the full Voice Controller on the same 24,000 events, including 9,600 turn endings and 14,400 events where the user should retain the floor. Thresholds are fixed before testing. Recall and false-positive rate include all events in their respective classes; UNCERTAIN decisions remain in the denominators and do not count as yields. Backchannel listening tests use the same clips and listeners against two separate baselines. One remains silent; the other uses random timing with matched response count and duration.

The full Controller combines 89.2% end-of-turn recall with a 6.7% false-positive rate, with decision coverage of 95.8% and conservative waiting on the rest. Listeners prefer its backchannels to both random timing and silence. Premature answers occur on 1.5% of 16,480 user turns, counted at most once per turn. Of 8,276 backchannels actually played, 0.7% inappropriately take the floor and force the user to stop. Both rates remain below their respective 2% and 1% limits. These results measure the balance we seek between responding when invited and leaving room for the person to continue.

The mobile optimization ablation measures how the waveform decoder, Slim Depth, and mixed precision contribute to the iPhone 17 Pro result. Together, they bring Edge’s median real-time factor from 0.77 to 0.53; Quality improves from 1.46 to 0.98. A real-time factor below one means speech is generated faster than playback. The matched warm, batch-one comparison holds text, voice, context, streaming chunks, and concurrent control load fixed.

Blind listening found no clear preference between the optimized and full-width, Mimi-decoded configurations. The optimized configuration received 48.1% of preferences for Edge (95% CI 44.7–51.6%; 1,641 judgments) and 49.2% for Quality (95% CI 45.8–52.6%; 1,660 judgments).

We test sustained on-device performance on MacBook Pro M5 Max, iPhone 17 Pro, and Galaxy S26 Ultra. Each device model and profile undergoes twelve 60-minute runs across three physical devices. We calculate p95 from complete-request RTFs within each run and report the median of those twelve p95s alongside median endpoint-to-audible latency. RTF measures speech generation only, with concurrent Voice control active. Peak memory covers the full enabled voice stack, including input audio, control models, caches, and buffers, and excludes concurrent PLM generation.

Edge is the default phone profile, with median per-run p95 real-time factor below one on all three device models and median endpoint-to-audible latency of 270 ms on iPhone and 300 ms on Galaxy. Quality is below the same RTF threshold on Mac; its median per-run p95 remains above one on both phones. These results support distinct deployment roles, with Edge for sustained mobile conversation and Quality for higher-fidelity speech on devices with more compute. Both retain the same context format and streaming interface, so applications can choose a profile to match their device budget.

The Future of Voice

KRORAINA 2 brings speech generation and interaction control into a shared real-time system. The next question is how that system can become a better collaborator through sustained conversation. We see four directions for this work.

• Longer conversational context. Compressed speech history and an editable delivery state could preserve continuity across longer exchanges without retaining every audio frame. With the person’s approval, the system could adapt to preferences for pace, pauses, and acknowledgment while preserving the chosen voice. A request such as “let me finish” should take effect immediately, overriding remembered habits. This is adaptation in how the system communicates, with preferences the person can inspect, change, or remove.

• Learning from interaction. We want to extend preference learning and reinforcement learning to the course of a conversation. Authorized interaction samples would let us compare timing, expression, and recovery after a correction while holding the task and intended content fixed. The objective is successful exchange, including whether the person had room to finish and whether the system adjusted when corrected. These signals can improve both generation and interaction policy, while the deterministic Voice Controller retains playback authority.

• Anticipatory generation. The architecture already allows preparation before permission to speak. We want to learn when anticipating a turn ending and preparing a likely continuation is worth the computation. If the person adds information that changes the answer, the preparation must be discarded. The research challenge is to reduce audible waiting without spending the phone’s compute and energy on work that will never be used. The complete input and current task state must still be checked before the Controller permits playback.

• More languages and ways of speaking. Beyond English and Spanish, we plan to extend coverage to French, German, Portuguese, Japanese, Korean, and Arabic. Each requires attention to pronunciation and prosody, as well as how speakers pause, acknowledge, overlap, and switch languages. Expansion should preserve intelligible speech, appropriate timing, safe interaction, and sustained local performance across languages.

People bring different languages, rhythms, and expectations to a conversation. Voice intelligence should learn to work with those differences, giving more people the freedom to question, correct, and direct AI in the way they naturally speak.

Smart Router

SADAIR Smart Router selects remote capability for each delegated task, balancing predicted quality with cost and time to a useful result. The resident Personal Language Model first decides whether delegation is needed. Tasks it can complete locally never invoke the Router.

Architecture

Smart Router V1 routes among three remote contracts. These are fast general, frontier general, and deep reasoning. Each contract specifies the model, system template, decoding policy, tools, data-handling terms, and the local integration that turns its output into a final response. The Router predicts how well each will complete the task with the information it is permitted to receive. Several candidates may be suitable, or all may be weak.

Permission is resolved before preference. The Context Compiler constructs each candidate’s view under deterministic policy, which may allow disclosure, require redaction or task-scoped confirmation, or deny it. A separate 11.1M-parameter local span detector can only tighten a field’s classification or return unknown. No model can grant itself permission. Sensitive disclosure requires the person’s authorization, and authentication secrets remain behind the local credential broker. Changes to authority or redaction invalidate the old predictions and rebuild every candidate view. Permission, destination, payload, and task revision are checked again immediately before anything is sent.

The quality predictor is a 117.5M-parameter, twelve-layer multilingual bidirectional encoder running on device in mixed W8. A deterministic packer fits the task and dialogue context, plus a descriptor of up to eight tokens for each candidate’s permitted view, within a 512-token budget. It predicts an independently normalized five-level quality distribution (π) for each contract. p is the probability of an acceptable final response, and μ is expected quality on a zero-to-one scale. Price, endpoint health, complete-contract latency, and effective throughput remain signed operational facts outside the network’s weights.

The deterministic Model Selector applies the chosen objective among eligible contracts. Intelligence maximizes μ, while Math and Coding use frozen domain profiles to resolve calibrated near-ties. Cost, Latency, and Throughput are optimized within an acceptable quality band defined by p. Auto adapts these priorities to live, foreground, or background work, with example decisions shown in the replay. Manual selection cannot override permission, budget, or compatibility. With no eligible contract, there is no dispatch. Task-scoped confirmation is requested only when authorization could restore eligibility.

Corrections and unresolved work can change which eligible candidate is likely to succeed. The Router cannot use these signals to initiate delegation, widen disclosure, or create action authority. Returned results are checked against the current task revision and integration contract, and stale results are rejected. The Router never speaks or updates the Memory System directly. Its complete local decision path has a p95 latency of 51.8 ms on a Galaxy S26 Ultra under the representative warm, batch-one, 200-token W8 workload.

Training and release evaluation

We trained the Router on all-candidate outcome matrices. Each candidate receives the same task under its permitted view, and the final result after local integration receives the label. This teaches the Router which candidates can succeed and how that changes when context is withheld. Executable checks score objective tasks. Anchored rubrics and blinded review score open-ended work, with 18,000 final artifacts each receiving three independent ratings. Correctness, instruction satisfaction, safety, and task completion remain hard gates.

The primary program contains 80,000 lineage-separated task contexts, with 60,000 for training and 20,000 for development, calibration, Selector fitting, and sealed evaluation. It schedules up to 375,000 candidate-outcome slots across these partitions. Tasks include rights-cleared and first-party material, audited synthetic and counterfactual examples, consented deidentified product-like tasks, and voice-derived requests. Counterfactuals hold the task fixed while varying the visible information and interaction state.

We fine-tune the encoder and quality head jointly, then calibrate the deployed W8 model per contract on held-out data. Quality bands and selection rules are fitted on a separate partition and frozen before evaluation. Price, latency, throughput, and Selector mode never enter the quality loss. A changed model or integration path requires new outcomes and recalibration. A separate paired-view diagnostic suite schedules 72,000 additional outcome slots using existing sealed contexts, testing how predictions respond when a candidate’s permitted information changes.

For live work, Smart Router reduces p95 dispatch-to-final-artifact latency from 9.18 to 7.06 seconds, a 23% reduction relative to the fixed Frontier contract, with similar observed acceptable-response rates. In foreground work, its cloud cost per acceptable final artifact is 6% lower than the same fixed contract. This gives applications a way to reduce waiting and remote spending while selecting for the quality a task requires.

The evaluation below compares the remote cost and waiting time required to deliver an acceptable final response. All policies use the same sealed outcome matrix and frozen final-artifact rubric. Comparisons include each fixed contract, cheapest eligible, static rules, a frozen embedding-and-MLP router, and an observed oracle that provides an upper bound on acceptable-response rate. Foreground results are plotted against cloud cost per acceptable final artifact. Live latency runs from local dispatch to the first complete, revision-valid, and locally integrated final artifact and is reported at p95. The intervals use paired bootstrap resampling by context lineage. Permission conformance, unauthorized-egress checks, and stale-result rejection remain separate release gates that a routing score cannot compensate for.

Compared with fixed Deep reasoning, the Router is within one percentage point of observed acceptance in both panels, at less than one third of its foreground cost per acceptable artifact and live p95 dispatch-to-final-artifact latency. Fixed Fast general and the frozen MLP router spend less in the foreground, with lower observed acceptable-response rates.

These routing gains remain inside the system’s authority boundary. Smart Router selects among eligible contracts; deterministic policy governs disclosure, and Policy & commit governs result acceptance, changes to the Memory System, and external effects.

Safety

Moving intelligence onto devices changes the safety boundary. Hosted systems can rely on monitoring, rate limits, rapid patching, and account-level intervention. Distributed models may continue working beyond those controls. We must decide which capabilities may leave our control, in which form, and with how much authority.

Privacy and safety are separate properties. A private model can still enable harm, while a model with modest general capability can become consequential when it gains access to the Memory System, tools, machines, or long-running tasks. We assess capability and authority together. Capability measures what a model can be made to do under strong elicitation, adversarial prompting, fine-tuning, and tool scaffolds. Authority measures the information it can access, the actions it can take, the people it can affect, and how long it can operate without review.

Release follows capability and authority

Each release undergoes a safety assessment of the complete configuration, including its quantization, context access, tools, and action policy. We test default behavior and maximum elicited capability across severe misuse, cyber operations, fraud, manipulation, autonomous tool use, deception, and attempts to undermine oversight. Prompt injection and tool-mediated instruction hijacking are tested separately. Third-party content can supply evidence; it cannot grant permission or weaken policy. Measured capability and granted authority determine the evidence and safeguards required for release. Neither compensates for a failed hard gate.

These boundaries are implemented in the Personal Intelligence Layer. The PLM emits typed proposals. Policy & commit checks authority, task revision, cancellation, and acceptance before results re-enter the live system, the Memory System is updated, or external effects proceed. Smart Router selects only among eligible contracts. Only the Voice Controller grants playback. No learned component can grant itself permission. These controls constrain authority; capability testing separately establishes what each model can be made to do.

We plan two model distribution paths; this article publishes neither model weights nor checkpoints. The three PLM Instruct configurations, both KRORAINA 2 profiles, and Smart Router are assigned to a resident deployment path using encrypted, signed, and versioned packages. We will release selected PLM Base checkpoints as open weights under Apache 2.0. Each checkpoint requires a separate safety case that remains valid without the ability to recall its weights. Encryption protects access to product weights; signatures verify package provenance and integrity. Neither establishes safe behavior. Base describes a training stage, not a safety class, and no checkpoint inherits the release decision of another configuration in its model family.

On-device describes where a model runs. Open-weight describes access to its weights. The license defines how those weights may be used, modified, and redistributed. Figure 16 maps four custody levels, from capabilities held in restricted environments (C0) and governed enterprise deployments (C1) to encrypted resident packages (C2) and open weights (C3). Open-weight release requires a higher pre-release standard because safety cannot depend on later recall, monitoring, or repair. Thresholds are maintained in a versioned release standard. New evidence or material changes can trigger reassessment; an updated standard does not rewrite the historical record of an earlier decision. When the evidence is insufficient, the release waits.

Customization carries responsibility

Every SADAIR release is evaluated against a common safety and rights baseline covering severe harm, exploitation of vulnerable people, coercive manipulation, abuse of third-party data, and unauthorized consequential action. Personal and enterprise policies operate within that baseline. A person may govern memory, disclosure, providers, initiative, and confirmation requirements. An organization may shape domain knowledge, workflows, permitted tools, and operational risk limits.

Every custom derivative becomes a new safety object. Changes to weights, quantization, tools, context access, authority, or release form require a new assessment and release decision. We repeat the relevant capability, misuse, and regression tests, recording the safeguards and remaining risk. Commercial ownership of a project does not grant release authority. Higher-risk deployments require approval from an internal release authority independent of the project’s commercial owner, with the power to delay or deny deployment. External red teaming and third-party evaluation are added in proportion to risk.

No model can be guaranteed never to produce harmful output. Release decisions depend on defined threats, measured capability, tested safeguards, and the remaining risk.

Responsibility continues after release

Responsibility continues through security updates, rollback paths, vulnerability reporting, and incident response. Versioned packages make affected configurations identifiable, while changes to models or their operating environment trigger reassessment. A distributed model’s safety case must account for attempts to remove refusals or combine it with stronger tools.

Learning from incidents must respect the same privacy boundary as ordinary use. Diagnostic sharing is opt-in, limited to a stated purpose, and subject to review before disclosure. Private local conversations are not collected by default. Personal-information filtering can reduce exposure, but identifiers can be missed and context can remain identifying after names are removed. Filtering does not establish anonymity or replace authorization. Safety feedback must improve the system without making private conversation the price of using it.

Personal intelligence also introduces risks that ordinary content evaluations miss. A system that remembers a person, speaks with them continuously, and responds to signals of frustration or vulnerability may become unusually persuasive. We test for sycophancy under correction, emotional manipulation, dependency, covert engagement optimization, and erosion of independent judgment. A stronger personal system should preserve the user’s ability to disagree, inspect what shaped an answer, leave the relationship, and retain useful capability without it. We will examine that requirement through fully unaided follow-up studies after assistance has been removed.

Ownership changes where responsibility is exercised, while responsibility itself remains. As intelligence moves closer to people, safety must travel with it. Evidence before distribution, bounded authority in use, and continuing accountability become part of every release.

What Intelligence Is For

Model capability enlarges the field of possible action. Purpose enters from lives models do not share. Families, laboratories, companies, and communities learn through experience what deserves effort, which risks to accept, and what must be preserved. General models can help people examine those judgments and act on them. The authority to make those judgments belongs to the people who bear their consequences.

The lasting gain should be a greater capacity to create. With better tools, people can investigate questions that once lay beyond their reach and develop methods others can build on. A discovery in one setting can become part of an organization’s practice or help a wider community solve a problem. Continuity gives that work a way to accumulate as the models improve.

A personal system can also erode the authority it is meant to preserve. A local system can still lock people into one provider. Memory can preserve error. A warm voice can manipulate. Remote delegation can improve an answer while moving control elsewhere. The design must preserve correction, reversibility, and bounded authority across the complete system. Local execution supplies reusable capacity at the edge and zero inference cost for work completed there. The Memory System and versioned controls provide continuity. Voice brings intelligence into live exchange, and Smart Router reaches stronger shared systems without widening the permissions already granted.

Making intelligence a lasting human capability is a system-level claim, so its unit of measurement must be the person and AI working together over time. Each study will compare the complete system with human-only, model-only, and the strongest simpler system available. The measures will cover useful work, correction, bounded authority, and the capability people retain after assistance ends. We will publish the methods, aggregate results, and audit evidence while protecting private histories and sealed tasks. A system that achieves more by concealing costs, weakening judgment, or enlarging its own authority has failed our mission.

Many builders will determine AI’s second history. Frontier laboratories will extend the common foundation, while people and institutions discover what intelligence becomes in use. SADAIR’s responsibility lies in the passage between those scales. We are building the models and systems through which people and organizations can put that progress to work and developers can create applications we would never have thought to build. What people learn through use should guide how those applications develop, so the capacity to shape intelligence grows with the capacity to use it.

When people can preserve distinct histories and choose what to contribute, knowledge can travel through tools, practices, evaluations, and discoveries without collapsing those histories into a single account of the world. A common foundation can then grow through many centers of judgment, while the capability formed through use remains with those who helped create it.

AI’s first history is measured in the capabilities of machines. Its second should be measured in the capabilities humanity keeps.