Chronological feed of everything captured from Riley Goodside.
It’ll be much worse at its worst but much better at its best
What am I missing? I didn’t give this post a ton of thought and I’m sorry if I’m misreading
Riley Goodside notes an hourly poll feature from his X feed that may function on iOS post-saving. He hypothesizes X's compression as the cause of potential issues. Plans to verify empirically.
Riley Goodside responds to a user note characterizing his X feed as an hourly poll. He expresses thanks and admits this alternative posting approach is superior to his original method. The exchange highlights iterative refinement in social media content strategy.
Riley Goodside targets the "spikes" in AI's jagged capability frontier, where isolated demonstrations reveal untapped potential. His goal is to craft posts that inspire readers to immediately apply novel ideas successfully, smoothing the frontier by expanding practical knowledge. This approach democratizes frontier-pushing via replicable insights.
ChatGPT's transcription feature automatically adds missing punctuation to text extracted from images, such as on a cake. This behavior was confirmed in a demonstration where the model corrected punctuation on three specific lines without indication. Users should account for such silent normalizations when relying on these transcriptions for precise work.
ChatGPT 5.5 Pro (Extended) autonomously simulates a chess game to test its reimplemented quiescence search algorithm, producing a multi-page PDF report with move analysis and a QR code linking to Lichess PGN viewer. The report critiques questionable moves, demonstrating integrated simulation, evaluation, and visualization capabilities. Generation took 33 minutes 26 seconds, though PDF formatting issues like empty bottom halves persisted.
ChatGPT 5.5 Pro (Extended) simulates a chess game to evaluate its reimplementation of quiescence search, producing a multi-page PDF report with analysis of questionable moves. The report includes a QR code linking to a Lichess PGN viewer for interactive game review. Generation took 33 minutes and 26 seconds, demonstrating advanced multimodal output capabilities despite minor formatting issues like empty space in portrait layout.
Requesting an LLM to create a reference image programmatically through code serves as an effective prompt technique to prevent mixels—unwanted artifacts in generated images. This method leverages the model's coding capabilities to produce a clean reference for guiding subsequent image generation. It functions as a reliable "incantation" applicable across various prompting scenarios.
Riley Goodside suggests using a smaller grid with borders preserved in initial output, followed by scaling code to target resolution. This approach warrants exploration for practical utility in graphics or UI rendering. It balances detail retention with efficient scaling.
ChatGPT 5.5 Pro (Extended) simulates a chess game to evaluate its reimplementation of quiescence search, an alpha-beta pruning technique for deeper search in tactical positions. It generates a multi-page PDF report analyzing questionable moves and includes a QR code linking to a Lichess PGN viewer for interactive review. This demonstrates advanced autonomous validation of chess engine improvements within a language model.
Riley Goodside's hourly poll on his X feed indicates initial positive reception to recent changes. He describes the experience as feeling incrementally better than before, though he likes it without rigorous comparative testing. This suggests early qualitative approval pending quantitative validation.
ChatGPT Images 2.0 generates amateur photos containing impossible self-referential elements, such as a man returning a canvas print to HomeGoods that depicts the exact candid photo being taken of him at that moment. This creates a logical paradox undermining the image's coherence. The example highlights limitations in the model's ability to handle temporal and causal consistency in generated visuals.
ChatGPT Images 2.0 (Pro) generated an amateur photo depicting a man returning a decorative canvas print to HomeGoods, where the print mirrors the exact candid photo being taken of him in that instant. This creates a paradoxical, nonsensical scene highlighting flaws in the model's image synthesis. The output reveals limitations in maintaining logical consistency during surreal prompt interpretations.
Charts in Riley Goodside's X feed are largely accurate except for a missing label at the 0.5 tick on the y-axis. The observer notes all other elements appear correct. This highlights a common visualization error in data presentation.
ChatGPT Images 2.0 (Pro) produces photorealistic images of cakes decorated with SVG code that, when extracted and rendered as a file, recreate identical cakes. This demonstrates recursive visual encoding where image generation models embed functional vector graphics into pixel outputs. The capability highlights advanced multimodal synthesis in proprietary models, enabling self-referential content loops.
Riley Goodside reports that Grok's Images 2.0 in Pro tier generated a high-quality image in one attempt, unlike the Thinking tier which required ~10 tries with unacceptable errors. This suggests a significant capability upgrade post-release. The difference is attributed directly to the tier and model version.
Riley Goodside's X feed processes chess games by searching for PGN notation, converting it to FEN, rendering as SVG, and feeding it as multimodal input to a final generation model. This pipeline enables visual chess analysis in generated outputs. The method is demonstrated in a ChatGPT session with detailed reasoning.
Riley Goodside observes that an updated game interface or character arrangement is less different than anticipated but includes targeted enhancements. Spell casters are repositioned to the backline behind frontline fighters, aligning with tactical formations. Characters adopt a more Final Fantasy-inspired aesthetic and are placed to the right of the boss for improved visibility and composition.
Python tools produce QR codes that serve as multimodal inputs to enhance LLM reasoning processes. These QRs are integrated into hourly polls on Riley Goodside's X feed. Detailed discussion available in thread replies.
An AI system identifies URLs within its processing context and automatically converts them into SVG representations of QR codes. In its reasoning summaries, these QR code SVGs are explicitly labeled as "helper images." This behavior demonstrates integrated multimodal output generation without explicit prompting.
Riley Goodside expresses an expansive view on AI capabilities, asserting that any task performable by a computer is now plausibly achievable with AI advancements. This reflects a paradigm shift where computational feasibility equates to AI realizability. The statement underscores accelerating progress in AI, diminishing prior boundaries on what systems can accomplish.
In a completely dark room, a candle's light source creates massively enlarged shadows that span a third or more of wall surfaces and extend to the ceiling. This geometry prevents recognizable silhouettes from forming against walls. The effect arises from the near-point light source relative to room dimensions, distorting projections unlike distant or collimated lighting.
ChatGPT's Images 2.0 model generates a visual of a standard six-sided game die where each face replaces numbers with scannable QR codes linking to Wikipedia articles. The QR codes are fully functional, demonstrating advanced integration of readable machine-readable data within photorealistic generative imagery. This showcases the model's capability for precise, utility-driven visual synthesis beyond mere aesthetics.
ChatGPT Images 2.0 (Pro) generates a cake visually depicting the polynomial f(x) = x^3 - 3x^2 + 2x. This demonstrates precise rendering of mathematical functions in image generation. The output aligns with the function's graph, showcasing advanced visual math comprehension in AI models.
ChatGPT Images 2.0 (Pro) produced a cake visually representing the cubic polynomial f(x)=x^3-3x^2+2x. A user nearly purchased it without scrutiny but was alerted by the accurate, literal depiction. This highlights AI image generators' capacity for precise mathematical visualizations that reveal intended product details.
I listened to that when I was 14 and I can still feel the cringe
Yeah their stuff’s confusing lately but honestly everyone in this industry has that issue
It’s post-slop. This is “Iron Man beats up Walter White” for people 1SD smarter.
Riley Goodside predicts a future where children perceive Pixar movies, created with computers but pre-AI techniques, as distinctly non-AI. This reflects evolving definitions of "AI" tied to modern generative models. The distinction will become incomprehensible to the next generation immersed in AI-native media.
Riley Goodside expresses nostalgia for the bonus content featured on DVDs, such as extras that enhanced the viewing experience. This reflects a perceived loss of tangible, value-added media features in the shift to digital streaming. The comment positions hourly X feed polls as a potential modern analog to such bonus materials.
Riley Goodside expresses pessimism about the AI development trajectory, stating that significant negative developments are still ahead. This hourly poll from his X feed captures a sentiment of ongoing decline rather than recovery. The brevity underscores a raw, unfiltered view likely tied to rapid AI advancements or hype cycles.
Riley Goodside announces his resignation from Google DeepMind effective today, citing personal health issues and focus on his daughter. He expresses gratitude for the opportunity provided by Logan Kilpatrick and honors the world-class talent at the organization. This marks the end of his tenure at GDM.
A user notes that Jensen Huang never said a particular statement, based on their search efforts. They interpret Goodside's post as intended satire but argue the satirical intent lacks sufficient clarity. This highlights challenges in distinguishing satire from factual claims on social media platforms like X.
Riley Goodside posits that large language models (LLMs) are not trained on chess due to its perceived lack of utility, rather than inherent incapability. He hypothesizes that deliberate post-training efforts would enable LLMs to play chess proficiently without relying on coding chess bots. This view stems from an hourly poll questioning the standalone importance of chess skill in LLMs.