Interactive Blog Posts

Explore how AI transforms each stage of the SDLC. Tune the cost equation to see when GenAI pays off, and toggle context sources to watch suggestions improve.

Toggle LayerNorm's projection and scaling on/off to see how it changes which keys attention can select. Visualize the "un-selectable key" problem in 2D.

Click nodes in side-by-side GAT vs GATv2 graphs. See how GAT's attention rankings stay static while GATv2's change dynamically with the query.

Step through GNN message-passing layers and watch information from distant nodes get squashed through bottleneck nodes. Add shortcut edges to see relief.

Write expressions with holes and examples in a RESL console. Watch synthesis candidates appear ranked by how well they match your specification.

Watch DAMP attack a code model by iteratively renaming variables. See the prediction shift in real-time, then toggle defenses on.

Make an edit in one location and see the model predict correlated edits elsewhere. Watch edit paths light up on the AST.

Compile and strip a C function, then watch NERO analyze the control flow graph to recover the original function name.

Pick an RNN architecture and see which formal languages it can recognize. Watch internal state evolve as strings are processed character by character.

See how AST paths provide richer context than flat tokens for code completion. Compare structural vs token-based model predictions side by side.

Enter code and see AST paths extracted, attention weights computed, and method names predicted. Explore how path-contexts capture code semantics.

Watch the decoder generate method name subtokens one at a time, attending to different AST path-contexts at each step.

Step through the weighted L* algorithm building an observation table and hypothesis automaton from membership queries.

Watch L* extract an interpretable DFA from a black-box RNN. See the observation table grow and the hypothesis automaton evolve step by step.

Adjust precision from 2-bit to infinite and watch how RNN counting ability degrades. See why practical RNNs are limited to regular languages.

Click pairs of AST leaf nodes to see the connecting path highlighted. Explore how paths capture syntactic patterns for property prediction.

Give examples to a synthesizer, then answer disambiguating questions to prune candidates. Watch the candidate count shrink to one correct program.

Provide one example test and watch the system generate a comprehensive test suite targeting edge cases and different code paths.

Compile and strip a C++ program, then watch the system reconstruct the class hierarchy from vtable patterns and constructor signatures.

Scan firmware for known CVEs using strand-based matching. See how FirmUp detects Heartbleed even across different CPU architectures.

Compare binaries from different compilers. Watch re-optimization bring them to canonical form, revealing their underlying similarity.

Move beyond concrete examples to abstract specifications like "any positive integer." See how abstract examples constrain synthesis more effectively.

Place membership queries on a 2D plane to learn a disjunction of predicates. Watch the hypothesis region converge to the target concept.

Build redundant extraction strategies for web data. Mutate the HTML and watch forgiving extractors survive while rigid ones break.

Step through a simulated video timeline and watch code being extracted frame by frame via region detection, OCR, and temporal diffing.

Compare code snippets using token, structural, and NL-based similarity. See how natural language descriptions bridge semantically equivalent but syntactically different code.

Align multiple product page DOMs to separate template from data. Reconstruct any original page from the extracted template and data table.

Extract instruction strands from binary functions and match them statistically. See how strand-based comparison survives cross-compiler differences.

Watch a model predict variable types from assembly usage patterns. See how FPU instructions signal floats and dereferences signal pointers.

Use one website's known structure to automatically synthesize an extractor for a different site in the same domain.

Watch data-driven refinement split imprecise abstractions into tight-fitting disjuncts on a 2D state space visualization.

Explore how natural language descriptions provide semantic signals for code similarity beyond syntactic structure.

Visualize store buffers under TSO and PSO memory models. See how abstraction reduces verification state space from 48 to 12 states.

Compare two program versions: see a big textual diff vs a precise semantic diff that shows only the behavioral change.

Extract tracelets from a binary's CFG, then search for matching functions in a target binary using tracelet matching.

Compare alphabetical vs statistically-ranked code completions. See how n-gram models put the right completion first.

Watch two threads operate on a BST simultaneously. See how logical ordering prevents conflicts without complex locking.

Watch memory operations get reordered by the processor, then see minimal fences synthesized to prevent buggy behaviors.