Personal AI News Digest
35 topics from 3 sources
4,018 words · ~20 min read
OpenAI released ChatGPT Work on July 9, 2026, an agent product for knowledge work that has crossed 10 million users in three weeks. Work consolidates ChatGPT, Codex, and cloud agents into a unified interface. On web and mobile, Work runs in the cloud; on desktop, it offers both cloud and local modes with full computer use. Each task gets a persistent working directory at /workspace/scratch where the agent can create files, install dependencies, and maintain state across sessions. However, continuity across tasks runs through ChatGPT's product layer rather than the computer itself: the agent accesses previous conversations via a Personal Context tool, files through the Library, and Projects as abstractions rather than filesystem directories. This design trades agent sovereignty for user safety and control. Work includes proactive task suggestions based on user context (calendar, email, preferences), scheduled automations that can run standalone or with heartbeat triggers to resume existing tasks, and browser automation through a remote Chrome service that maintains persistent login sessions and a permission ledger. The agent can click, type, scroll, take screenshots, and move files between browser and workspace, though cloud-based constraints prevent some tasks that work locally (e.g., Amazon and Google Photos rejected the cloud browser). Work integrates plugins from the Plugin Directory (1,000+ available), which can contain apps, skills, and custom UI, but lacks a discovery layer—the agent doesn't suggest relevant uninstalled plugins even when they would be more efficient than web search. Work and Chat will merge by year-end, making these design choices the default for ChatGPT's billion weekly users. Key tensions remain: whether to expose local task state across cloud, how to coordinate between tasks without a meta-layer agent, and how to improve plugin discovery.
Sources Latent.Space
Links Unpacking ChatGPT Work: the Agent for a Billion Users, ChatGPT Work, a busy launch: three new models, crossed 10 million users, Greg Brockman has confirmed that they will merge by the end of the year, cloud computer is persistent too, working directory under, Personal Context, Library, Projects, Scheduled Tasks, inspect the page, click, type, scroll, take screenshots, manage tabs and dialog…, persistent profile, a small permission ledger is synchronised into its computer alongside the works…, July 9 launch, create personal plugins, asked it to search for flights
Alibaba announced Qwen3.8-Max, a 2.4T-parameter sparse MoE model with open weights releasing next week, alongside Qwen3.8-27B. API pricing: $2.00/M input, $6.00/M output, $0.25/M cached tokens. The model activates ~4% of parameters per token (95B active / 2.4T total). Key capabilities include 10+ days of autonomous coding, 500+ turns of chip design optimization, 365-day e-commerce simulations, and native multimodal reasoning. Third-party evaluations: Frontend Code Arena #4 (1,668 Elo), Vision Arena #2 (1,305), Vals Index #2 among open-weight models (66.1 overall, #10 of 43). Benchmark scores: SWE-bench 87.3% (vs GPT-5.5 82.6%, GLM-5.2 83.3%, Claude Opus 4.8 89.2%), Terminal-Bench 2.1 67.4 (up from Qwen 3.7 Max's 61.0). Vals reported cost-per-test $2.68 vs Claude Opus 4.7's $6.17 at matched performance (66.1 vs 66.1). Context window 1M, max output 128k. Price cut from prior Qwen Max ($2.50/$7.50 to $2.00/$6.00 input/output). Qwen 3.7 Max scored 57.5 on Vals Index; 3.8 Max's 66.1 represents 8.6-point gain in ~2.5 months. Infrastructure reality: requires 8+ H100/B300 GPUs minimum; not a local-inference commodity model. License ambiguity flagged: some read terms as restricting use in USA, EU, UK, Korea; Alibaba did not clarify in-thread. Strategic significance: shift from API-only Max models to open-weight flagship, driven by competitive pressure from DeepSeek, Kimi, GLM in Chinese open-weight frontier.
Sources AINews
Links Qwen 3.8 Max, @Alibaba_Qwen, @Alibaba_Qwen, @Alibaba_Qwen, @ValsAI, @arena, @arena, @ValsAI, @ValsAI, @ValsAI, @ostrisai
Alongside Qwen3.8-Max, Alibaba announced Qwen3.8-27B will also go open-weight. Multiple observers highlighted this as potentially more consequential than the flagship for practical deployment, as it is the tier most likely to become usable across open-source stacks and local-serving ecosystems. Developers framed the release in terms of downstream harnesses and agents (Hermes Agent, Command Code, Baseten) that can slot it into existing OpenAI/Anthropic-compatible workflows. One interpretation suggested Qwen 3.8 Max may be exceptionally strong on image recognition and potentially distillable into the 27B variant for task-specific parity, creating a route from flagship capability to laptop-deployable specializations.
Sources AINews
Links @kimmonismus, @TheZachMueller, @Alibaba_Qwen, @teortaxesTex
Jamin Ball and others highlighted that frontier open models like Qwen3.8-Max, Kimi K3, and GLM-5.2 are operationally open but not broadly accessible for local inference. Qwen3.8-Max >2T params requires at least 8 H100/B200 GPUs; Kimi K3 (~104B active per token) needs >1TB memory and 64+ accelerators in supernode setups. Token efficiency matters more than vanilla token prices: these models are enormous and expensive to serve. StableQuan noted long, RAM-heavy prompts and slow tool calls make giant models painful on consumer hardware, recommending API use instead. The practical competitive effect may be less dramatic than leaderboard headlines suggest, since deployment requires careful harness engineering and significant infrastructure investment.
Sources AINews
Links @jaminball, @stablequan
If the 95B active / 2.4T total parameter count is accurate, Qwen3.8-Max activates approximately 4% of total parameters per token, compared to Qwen3-235B-A22B which activates closer to 10%. This sparse routing strategy allows Alibaba to cut API pricing while scaling total parameter count upward: bigger expert pool, lower active footprint, lower effective inference cost, assuming routing and systems optimizations hold up in production. The architecture reflects a broader trend in frontier open models toward sparse MoE designs that look huge to store yet remain cheap to run.
Sources AINews
Links @ZhihuFrontier, @ProfTomYeh
Long-horizon capability is a model × harness property, not just a model property. Qwen3.8-Max was pitched less as a chatbot and more as a substrate for long-running work, with emphasis on 'coding and cowork' rather than generic assistant use. Cline reported that many open models are RL-trained to spend extra tokens on verification (rerunning tests, checking builds, rereading diffs) and deliberately letting them 'work how they were trained to work' yields ~20% gains from harness changes alone. Hermes Agent, Command Code, and other frameworks are shipping with voice activation, plugin/API expansions, and token-efficiency improvements. A paper categorized 41 agent failure modes by interaction edges between model, user, harness, tools, memory, and environment rather than blaming a single component. Cloudflare launched @cloudflare/computer, an agent runtime that dynamically routes between isolates and Linux containers. Cursor reported 20–30% better token efficiency for cloud agents and 80% better efficiency on computer-use runs.
Sources AINews
Links @cline, @ZhihuFrontier, @Cloudflare, @cursor_ai, @cursor_ai, @hwchase17, @omarsar0
Multiple observers noted that Chinese labs now dominate the open-weight frontier. Kimi K3, Qwen3.8-Max, GLM-5.2, DeepSeek, and MiniMax define much of the open frontier, while US labs retain lead positions mainly in select closed offerings. Artificial Analysis data cited in-thread indicated Chinese frontier models historically trail top US models by 3–9 months, while the open-weight frontier itself has been dominated by Chinese labs for roughly two years. Top three spots in front-end design are now shared between two Chinese and GLM models. Some posters argued Chinese aggregate research capability may already exceed US labs despite resource asymmetries. The moat may be shifting from raw pretraining to post-training, agent harnesses, inference infra, distillation pipelines, and developer lock-in.
Sources AINews
Links @kimmonismus, @matvelloso, @_micah_h, @kimmonismus, @teortaxesTex
Kimi K3 continued to post strong downstream results. RSIBench-Data reported Kimi K3 + Kimi Code at 27.317% weighted score across six automated-research benchmarks, including 50% SWE-bench Verified and 17% SWE-bench Pro. However, inference provider materially changed leaderboard outcomes: one provider produced degraded looping behavior while Modal's endpoint yielded #1 results on CEO-Bench. On the serving side, Databricks claimed #1 Kimi K3 inference speed/latency on Artificial Analysis, quoting 239 tok/s, with separate single-node numbers from Casper Hansen of 947 tok/s batch-32 decode and 152 tok/s single-user on a single B300 node. This underscores that harness and endpoint choice are first-class engineering variables.
Sources AINews
Links @tonychenxyz, @FanqingMengAI, @Yuchenj_UW, @casper_hansen_
DeepSeek V4 Flash (0731) emerged as the day's strongest cost-adjusted agent model story. Reported 57.1% / 63.0% on WeirdML for Flash-0731 high/max. Vals called DeepSeek V4 Flash (0731) the cheapest model on the Vals Index above 60, and 35× cheaper than the next best model at that threshold, with most of the advantage coming from coding and agentic tasks. Together AI immediately positioned it as a production endpoint for long-running agents.
Sources AINews
Links @htihle, Vals, Together AI
Intology claimed its Locus automated AI research system is SOTA on PostTrainBench. Locus-post-trained Qwen3 1.7B Base variants surpassed the official human post-trained Qwen3 1.7B Instruct release under expanded compute budgets. On live Kaggle competitions, Locus reached the 4th highest average rank after 16 days. This demonstrates that automated post-training pipelines can now match or exceed human-tuned baselines at scale.
Sources AINews
Links @intology
MiniMax H3 is a general-purpose multimodal generation model supporting text, image, video, and audio in a single context. Arena ranked it #1 open model in Video Arena across both text-to-video and image-to-video, with +280 points over the next-best open model; in image-to-video it was effectively tied for #1 overall. MiniMax claimed H3 is SOTA open video generation on both Arena and Artificial Analysis benchmarks. H3 is usable on RTX 5090-class hardware with ComfyUI stack around 40GB and ~5.5 min for 5s generations in early tests. However, context orchestration, 2K regeneration, and sparse attention remain server-side or otherwise restricted. License ambiguity: initial reading suggested geography restrictions (US/EU/UK/Korea), but clarification indicated those regions require formal authorization rather than outright prohibition.
Sources AINews
Links @arena, @MiniMax_AI, @ComfyUI, @victormustar, @MiniMax_AI, @VictorSuOrtiz
TokTier, a stateful tokenization service, reuses and repairs tokenized prefixes for agent sessions, reporting 16–34% TTFT reductions under vLLM and up to 437× speedups over standard Hugging Face tokenization in incremental repair scenarios. Tokenization can consume up to 64% of TTFT in cached-agent workloads. This is a first-class bottleneck once agent transcripts get long and cache hit rates are high.
Sources AINews
Links @omarsar0
OpenAI detailed a new GPT-Live architecture supporting full-duplex conversation—listening while speaking—by separating a dedicated fast audio path from slower asynchronous reasoning/tool-use paths. Session startup reduced from six network round trips to one. The system supports async compaction for long-context voice sessions. This represents a significant redesign of the realtime voice stack to enable continuous interaction.
Sources AINews
Vikhyat announced Photon 2.0, a compiler that turns models like Moondream, Qwen 3.5, and Gemma 4 into megakernels representing the whole forward pass as a single GPU program. The compiler uses a tracer DSL for dataflow specification and a CPU cost model to prune scheduling candidates before compilation. This reflects a broader trend where classical hand-optimized GPU kernel work is being progressively automated and commoditized.
Sources AINews
Links @vikhyatk, @waterloo_intern
Jina released jina-reranker-v3.5, a 0.6B listwise reranker scoring 63.20 nDCG@10 on BEIR, beating Qwen3-Reranker-4B at roughly 7× fewer parameters. This demonstrates efficient reranking at scale.
Sources AINews
Links @JinaAI_
Cloudflare launched @cloudflare/computer, an agent runtime that dynamically routes between lightweight isolates and full Linux containers so each agent gets 'a computer of its own.' This enables better isolation and resource management for long-running agent workloads.
Sources AINews
Links @Cloudflare
Cursor reported 20–30% better token efficiency for cloud agents and 80% better efficiency on computer-use runs. The company launched plugins for Google Workspace access across Gmail, Drive, Calendar, Docs, and Sheets, moving coding agents closer to general work automation.
Sources AINews
Links @cursor_ai, @cursor_ai
LangChain signaled Managed Deep Agents moving to public beta, with built-in evals, memory, OAuth tool access, channel integrations, and sandboxing. This represents a managed platform for long-horizon agent deployment.
Sources AINews
Links @hwchase17
OpenAI announced an internal next-generation model found 10 new results on long-standing open problems in mathematics and theoretical computer science for roughly $2,000 in token cost at GPT-5.6 Sol rates. This prompted both excitement about AI research capability and skepticism about total attempt cost vs solved-cost accounting.
Sources AINews
Links @OpenAI, @NickEMoran
Epoch reported roughly 2,500 high/critical CVEs disclosed in July across 21 major tech organizations, about 5× the prior monthly record before Anthropic's autonomous vulnerability-finding disclosure. This reflects increased security research activity and disclosure velocity.
Sources AINews
Links @EpochAIResearch
The White House reportedly invited OpenAI, Anthropic, Google, and Meta to review a new voluntary AI framework and finalized new cybersecurity tests and hacking benchmarks. This signals continued government engagement on AI safety and security standards.
Sources AINews
Links @steph_palazzolo, @AndrewCurran_
Google rolled out Gemini Spark auto-browse using Chrome to act in logged-in accounts for errands with user confirmation on sensitive steps. This extends agent capabilities to web automation with user oversight.
Sources AINews
Links @Google
Sakana launched Namazu API, described as a Japanese-focused LLM built on Kimi and tuned for Japanese language, culture, and business contexts, with reduced unnecessary refusals and bias. This represents localization of frontier models for specific markets.
Sources AINews
Links @SakanaAILabs, @SakanaAILabs
LiteParse added direct structured PDF extraction for form fields, checkbox states, annotations, embedded images, vector graphics, tagged structure, and word-level bounding boxes in milliseconds per page for simple pages. This reduces reliance on vision models for document processing.
Sources AINews
Links @llama_index
The Hermes Agent ecosystem shipped a substantial 'Herald' release with voice chats, plugin-based desktop features, A2A protocol, outbound webhooks, research and productivity skills, and token-efficiency improvements. This represents maturation of the open agent framework.
Sources AINews
Links @Teknium
Zero-Mem removes LLM calls from memory maintenance and only invokes an LLM at final answer time, cutting memory-operation cost by 57.6% versus the fastest baseline at matched budget. This reflects a trend of de-LLM-ifying components where possible.
Sources AINews
Links @dair_ai
Shahules argued benchmarks should release full trajectories, not just scores, because task defects and brittle verifiers can dominate failures. RSIBench-Data and other research-automation benchmarks are increasingly measuring specialized research harnesses, not just model intelligence. This highlights the importance of understanding whether failures stem from weak models, brittle verifiers, or under-specified tasks.
Sources AINews
Links @Shahules786
Ddkang reported >9% lower MATH accuracy when constructing noisy-data training more rigorously, pushing back on claims that RLVR with 100% noisy data matches clean-data training. Separately, ArmenAgha shared a result where optimizing a proxy objective improved selected velocity MSE but made actual rollout inference worse on held-out data. These findings underscore that 'self-improvement' headlines still collapse if evaluation is not robust.
Sources AINews
Links @ddkang, @ArmenAgha
Multiple posts suggested GLM-5.3 is imminent. GLM-5.2 Max already ranks #2 overall and #1 open-weight model in Frontend Code Arena, maintaining strong competitive position in the Chinese open-weight frontier.
Sources AINews
Links @AiBattle_, @arena
A comprehensive two-hour deep dive into modern inference engineering covering the full stack of techniques for turning open model weights into fast, reliable, production APIs. The conversation spans cache-aware routing, disaggregated prefill/decode, quantization, speculative decoding, KV-cache movement, model parallelism, GPU kernels, and the race to achieve 10× speedups on frontier models.
Key technical findings: In one GLM-5.2 experiment, quantizing more of the model actually preserved benchmark quality while increasing throughput by 20%, because quantization errors in different layers canceled each other out. Baseten's research showed 20% better quantized GLM-5.2 than NVIDIA by predicting which layers have canceling errors using KL divergence. Typical unoptimized inference on a 1T-parameter model runs at 30–50 tokens/second; stacking optimizations (NVFP4 quantization ~30–40% speedup, speculative decoding ~2×, disaggregation ~2×, better kernels) can reach 300–400 tokens/second or more, though 4–6× improvements are more common than the aggressive 10× target.
On model support: launching a production-ready API for a new open model (e.g., GLM-5.2, Kimi K3) requires far more than just running it through vLLM. Work includes redoing quantization calibration, training traffic-specific speculative decoders on representative datasets (coding, agentic use cases), building support for novel architectures (GLM-5.2's sparse attention from DeepSeek), and extensive testing. Baseten retrofitted Kimi's vision encoder onto GLM-5.2 by training only a small projector layer while freezing both the encoder and language model weights, achieving 56% on MMLU Pro without degrading GLM-5.2's base quality. Loop detection and race conditions in inference kernels remain production issues; some models (GLM-5.2, DSV4) collapse into repeating tokens, often due to subtle kernel synchronization bugs or hardware interconnect timing rather than model weights.
On quantization and quality: most inference optimizations are lossless (KV caching, speculation). Quantization is the main lossy step. Baseten's approach selects which layers to quantize by predicting error cancellation, achieving better fidelity than uniform quantization. Kimi publishes vendor benchmarks to hold providers accountable for quality preservation.
On hardware and parallelism: tensor parallelism (TP) requires high-bandwidth interconnects like NVLink and is good for latency; expert parallelism (EP) works well for MoE models and reduces communication overhead. Pipeline parallelism is rarely used except for multi-node inference on models larger than single-node HBM. B200 GPUs have 180 GB HBM per GPU (1.44 TB per 8-GPU node in FP4); H100s have 80 GB. Rubin (next-gen NVIDIA GPU) will emphasize systems-level improvements: faster memory bandwidth, better CPU-GPU interconnect, and KV-cache-aware routing. Mega kernels are bearish; specialized tensor cores and TMA (tensor memory accelerators) in newer GPUs reduce the need for hand-fused kernels. ASICs for AI are viable for vertically integrated model labs (e.g., OpenAI + custom chip) but face obsolescence risk as model architectures evolve; GPUs are trending toward more specialization (systolic arrays, tensor cores) without burning weights into silicon.
On video and diffusion: video generation faces a quadratic attention bottleneck. A 5-second 480p video at 16 FPS compresses to ~35,000 tokens in latent space; full attention is O(n²), making long-form video infeasible. Solutions: sparse attention (top-K, spatial/temporal locality) degrades quality; full attention requires enormous compute. Autoregressive video generation could scale better (enabling streaming, applying LLM optimizations like spec-dec) but current open-source autoregressive models (e.g., Wan 2.2 vs. closed Kling/Veo) have poor quality. Diffusion models attend bidirectionally and can fix past frames for consistency, but require keeping all tokens in memory. The future likely involves hybrid autoregressive-diffusion approaches. Audio and speech have largely moved to autoregressive (LLM-based waveform generation), while image generation uses both autoregressive (Nano Banana, GPT Image) and diffusion (Diffusion Gemma, 25B). Text diffusion remains underexplored despite potential for coherence across long sequences (e.g., storyboarding movies).
On training-inference convergence: inference optimizations (quantization, speculation) increasingly require training-aware techniques. Quantization-aware training, distillation, and post-training are needed to preserve model quality after NVFP4 quantization. Inference for training: RL rollouts bottleneck on slow inference engines; faster inference accelerates training loops. Models are beginning to optimize their own inference: GLM-5.2 was used to write GPU kernels for its own inference engine, analyzing profiling traces and generating optimized CUDA code in a closed loop.
On continual learning: two paths exist—pushing new knowledge into weights (only fixes one-hop facts, breaks multi-hop reasoning) or using KV-cache compaction (e.g., STILL paper) to maintain an infinite, compacted context. KV-cache compaction preserves reasoning chains and is the preferred direction; inference changes minimally (just update the KV cache).
Future trends: models will grow larger, hardware more powerful, and demand for 10–100× more tokens will drive infrastructure scaling. The biggest bottleneck is network speed: KV-cache transfers between nodes currently require HBM staging (4.5 TB/s) before NIC transfer (~100 GB/s), a 50× gap. Faster NICs could theoretically enable 100× speedup in disaggregated serving. New modalities (Cosmos world models, speech-to-speech, multimodal agents) will expand the inference engineering surface. Inference engineering is shifting from a CUDA kernel problem to a systems and infrastructure problem, with emphasis on KV-aware routing, disaggregation, and orchestration across clusters.
Sources Latent.Space
Links The Inference Engineering Mas…, The Inference Engineering Masterclass — Philip Kiely & Ali Taha, Baseten, covered Baseten, monster $13B round, AI Infra decacorns, the Inference Inflection, Open Weights debate, Kimi K3, spoken at AI Engineer, definitive book on Inference Engineering, How to build a day-0 API for Kimi K3, 22580: From GPT2 to Kimi3, Explained, https://www.linkedin.com/in/philipkiely, https://x.com/philipkiely, https://www.linkedin.com/in/aliestaha/, https://x.com/waterloointern
Researchers from University of Toronto, Vector Institute, University of Cambridge, and ServiceNow have built a proof-of-concept computer virus that uses open-weight LLMs running on compromised GPU nodes to autonomously detect vulnerabilities, devise tailored attacks, and self-replicate across networks. The worm uses a reasoning graph with specialized nodes (Plan, Judge, Action, Summary, Progress, and others) to decompose attack reasoning and avoid context bloat. It achieves ~80% success on vulnerability detection, ~53% on exploitation, and 88% on self-replication, yielding ~37% overall attack success rate. The agent operates fully decentralized with no single point of control, continuously self-replicating into a swarm where different replicas retry difficult hosts with fresh reasoning trajectories. The researchers used only an open-weight LLM published in 2025 that fits on a single A100 GPU with 80GB VRAM, with no reliance on monitored vendor APIs. This demonstrates that self-sustaining AI-driven cyber-threats are no longer theoretical and suggests the future internet will resemble a complex ecology of attacker and defender AI agents.
Sources Jack Clark from Import AI
Dwarkesh Patel argues that as AI models become smarter and better monetize compute, GPU rental prices will rise significantly. He estimates that if an H100-equivalent system could run a true human-level software engineer, at current market rates for software engineers that H100 should rent for over $250k per year—15x today's spot prices. This price increase will occur because AI is currently cheap relative to human labor partly due to capability gaps; as those gaps close, low-value applications like short-form video generation will be priced out. Patel expects this to be temporary; massive roboticization of the compute supply chain should eventually bring prices down closer to raw input costs, though by that point civilization will be deep into the singularity.
Sources Jack Clark from Import AI
Links Why compute might get 10x+ more expensive in coming years (Dwarkesh Patel, subs…
Senior leaders from OpenAI, Anthropic, Google DeepMind, Thinking Machines, Meta, and Safe Superintelligence Inc.—including chief scientists, cofounders, and CEOs—have issued a statement requesting that the US government support an international effort to develop technical and governance tools to deliberately pace frontier AI development. The signatories acknowledge that AI companies are close to automating AI research and that capability development risks accelerating beyond human ability to understand or control resulting systems. They note that each company and country faces intense competitive pressure not to unilaterally slow acceleration, and the world currently lacks the tools to deliberately pace frontier-wide progress. The statement frames this as a collective action problem requiring coordination among companies and governments to give society time to address emerging risks and develop security measures.
Sources Jack Clark from Import AI
Researchers from Princeton, Cornflower Labs, UK AI Security Institute, University of Toronto, UC Berkeley, Georgetown, Johns Hopkins, Golden Gate Institute for AI, AI Digest, and Stanford conducted a 'shadow evaluation' where Claude Opus 4.8 running in the OpenClaw harness attempted to solve two unpublished NeurIPS 2026 research papers—one on LLM persona structure and controllability, another on distribution shift detection for tabular foundation models. The paper authors graded the agent's output as conference submissions. Both attempts were rejected: the Personas paper scored 2 ('Reject') and the TabPFN paper scored 1 ('Strong Reject'). Reviewers cited poorly motivated data and experiments, no novel contribution, and impenetrable prose. While agents solved the engineering problems necessary to conduct research, they failed to produce original research at top-conference caliber, committing early to narrow research paths, ignoring synthetic feedback, and struggling to reverse course from unpromising approaches. This suggests that despite extraordinary engineering capability, today's AI systems lack the intuitive, tasteful creativity needed for paradigm-shifting research insights, potentially delaying timelines for fully automated AI development.
Sources Jack Clark from Import AI
Links Can AI agents conduct open-ended AI research? Early evidence from two case stud…
OpenAI used an internal version of Astra, its next major AI model, to solve ten open problems spanning high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics. These problems are of substantial interest to their respective mathematical communities and several are of broad interest across mathematics. Henry Yuen, associate professor of computer science at Columbia University, emphasized the importance of the solutions, noting he deeply cares about the answers to these questions. The results demonstrate that AI systems can now reliably drive forward the frontier in domains like math and theoretical computer science where solutions are easy to verify, though it remains unclear whether AI systems can generate their own novel research questions to expand human knowledge.
Sources Jack Clark from Import AI
Links Ten advances in mathematics and theoretical computer science (OpenAI blog)