Company Profile
Unlock undiscovered discovery,
expand tech innovation with reliable AI.
With the mission of “Unlock undiscovered discovery, expand tech innovation with reliable AI,” Apex Intelligence is committed to pioneering tech research foundation models, with the approach of RSI.
Our goal is to transform AI from a research copilot into a research co-worker by developing self-evolving foundation models that accelerate technological discovery and scientific exploration and, over time, match and ultimately surpass the collective research capabilities of 100,000 leading AI scientists across disciplines, forming a scalable “AI Researcher Network.”
We are steadily advancing toward this vision. This is not merely about improving efficiency, but about fundamentally reshaping the R&D paradigm.
This approach applies to fields where research workflows can be structured and outcomes measured, including chip design and simulation, molecular discovery, and quantitative strategy iteration. These fields share three defining characteristics: high value, intensive reliance on human expertise, and high experimentation costs.
Core Positioning
Research is the engine of recursive self-improvement.
Explore milestones & sources
Selected research milestones, shown schematically. Curve height and spacing do not compare performance across fields. Hover near a curve or tap to explore its milestones.
Chip Design
1973-04 · technical report
SPICE: Simulation Program with Integrated Circuit Emphasis
SPICE combined nonlinear DC, small-signal, and transient analysis in a general-purpose circuit simulation program. It made numerical analysis of circuit behavior a reusable step before fabrication.
Primary source ↗1980 · book publication
Introduction to VLSI Systems
Mead and Conway brought VLSI design abstractions, design rules, and system-level methods into a shared reference. The book helped make complex chip design teachable as a structured workflow.
Primary source ↗1983-12 · technical report
Magic: A VLSI Layout System
Magic embedded design-rule checking, connectivity, interactive stretching, and routing into an editable layout system. It joined layout creation and continuous checking in one environment.
Primary source ↗2013 · conference paper
Yosys: A Free Verilog Synthesis Suite
Yosys provided a free, open-source Verilog synthesis framework with an extensible RTL-to-netlist flow. The project made synthesis infrastructure more inspectable and reusable for engineering and research.
Primary source ↗2019 · conference paper
OpenROAD: Toward a Self-Driving, Open-Source Digital Layout Implementation Tool Chain
OpenROAD presented an open tool chain for automated digital layout implementation. It connected stages of physical design in a shared flow without implying that every chip-design decision was autonomous.
Primary source ↗2020-04-22 · research preprint
Chip Placement with Deep Reinforcement Learning
The paper framed macro placement as a reinforcement-learning problem over chip netlists and canvases. Its learned policy transferred experience across blocks while remaining one stage within a broader design flow.
Primary source ↗
Robotics
1985-09-01 · MIT AI working paper
A Robust Layered Control System for a Mobile Robot
Brooks described asynchronous behavior layers in which higher-level behaviors could suppress lower-level outputs while the lower layers continued to run. The architecture offered a situated, reactive approach to mobile-robot control.
Primary source ↗2005-10-08 · autonomous field competition
Stanley Wins the DARPA Grand Challenge
Stanford's Stanley won the second DARPA Grand Challenge on a 132-mile desert course. The event demonstrated a vehicle-scale autonomous stack operating through an extended real-world course.
Primary source ↗2017 · conference publication
Learning Hand-Eye Coordination for Robotic Grasping with Deep Learning and Large-Scale Data Collection
A learned visual model continuously updated gripper motion from monocular images during grasping. The work joined perception, action, correction, and large-scale robot data in a closed feedback loop.
Primary source ↗2018 · conference publication
QT-Opt: Scalable Deep Reinforcement Learning for Vision-Based Robotic Manipulation
QT-Opt learned closed-loop grasping policies from large-scale real-world interaction. The robot updated its strategy from new visual observations instead of executing a single fixed grasp plan.
Primary source ↗2022-04-04 · research preprint
Do As I Can, Not As I Say: Grounding Language in Robotic Affordances
SayCan combined language-model suggestions with affordance values that estimated which learned robot skills were feasible in the current setting. It connected semantic task planning to physically grounded skill selection.
Primary source ↗2022-12-13 · research preprint
RT-1: Robotics Transformer for Real-World Control at Scale
RT-1 represented camera images, language instructions, and robot actions within a transformer policy trained on diverse real-world episodes. The study tested how a shared model could scale across many robot tasks.
Primary source ↗2023-07-28 · research preprint
RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control
RT-2 co-fine-tuned vision-language models on web-scale tasks and robot trajectories, expressing actions as tokens. The research examined how semantic representations could transfer into closed-loop robotic control.
Primary source ↗
AI
2017-06-12 · research preprint
Attention Is All You Need
The paper introduced the Transformer, an encoder-decoder architecture based on attention rather than recurrence or convolution. Its parallelizable design became a foundation for later large language and multimodal models.
Primary source ↗2020-05-28 · research preprint
Language Models are Few-Shot Learners
The GPT-3 study showed that scaling an autoregressive language model improved task performance from instructions and a few text examples without task-specific gradient updates. It also documented important limitations and methodological concerns.
Primary source ↗2022-03-04 · research preprint
Training Language Models to Follow Instructions with Human Feedback
The InstructGPT work combined supervised demonstrations, ranked model outputs, and reinforcement learning from human feedback. It studied how post-training could better align language-model behavior with user intent.
Primary source ↗2022-10-06 · research preprint
ReAct: Synergizing Reasoning and Acting in Language Models
ReAct interleaved language-model reasoning traces with task-specific actions. The method let models update plans while interacting with external information sources or environments.
Primary source ↗2024-08-12 · research preprint
The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery
The paper presented a pipeline that generated research ideas, wrote and ran code, summarized experiments, drafted papers, and performed simulated review in selected machine-learning settings. The authors also reported implementation and evaluation limitations, so the result is a research prototype rather than evidence of general scientific autonomy.
Primary source ↗
What we see
Human technological progress often begins when scientific breakthroughs open new S-curves, then research discovery and engineering turn them into industries. Today AI improves engineering and execution, while scientific question framing, experimental-path selection, result judgment, and breakthroughs at the knowledge frontier still depend heavily on human researchers.
What we believe
Research is not an application add-on; it is the engine of self-evolution. The next AI revolution is not only about more parameters or data. Models must accumulate experience, receive objective feedback, and actively discover.
Our answer
We do not limit the model to an “AI assistant” that answers questions. We train AI researchers to participate in research, execute it, and improve its methods—going beyond restating known human knowledge to autonomously explore unknown laws through verifiable experiments and objective feedback.
Our Team
Apex Intelligence brings together a team with industry experience at Google, Microsoft, Amazon, Kimi, Zhipu AI, and Ant Group, as well as research backgrounds from Tsinghua, PKU, HKU, Harvard, MIT, UMich, Oxford, and UCL, among others. Our members have worked on large-scale AI systems, frontier model development, and rigorous academic research, combining practical engineering execution with deep technical expertise.
The team's work spans many of the core areas needed for self evolving AI, including large scale foundation models, language and vision language reasoning, reinforcement learning, agentic systems and tool use, multi agent coordination, AI safety and alignment, robotics, and AI for science.
Academic Backgrounds
Tsinghua
PKU
HKU
Harvard
MIT
UMich
Oxford
UCL
Join Us
At Apex Intelligence, founding members are entrusted with the company's most fundamental and technically critical challenges. You will help build a self-evolving ASI system from zero to one, taking part in the formation of a new frontier in intelligence.
We are now recruiting founding core members across all disciplines, with internship pathways open for all roles.