🤖💻 AI Daily Digest — August 1, 2026
OpenAI published a rare technical deep-dive on July 29 explaining why GPT-5.6 Sol looked weak on ARC-AGI-3, a benchmark where agents explore unfamiliar 2D puzzle games and infer the rules through trial and error. With the official evaluation harness, Sol scored just 7.8%, and GPT-5.5 could barely play at all (0.4%). The team discovered the problem was not the model but the harness: after every game action, all private reasoning was discarded, and a rolling truncation window deleted older actions as history grew. Sol was effectively being asked to figure out each game anew on every single move.
Enabling two Responses API settings — retained reasoning and context compaction — tripled the score on the public task set from 13.3% to 38.3% while cutting output tokens by roughly 6x. With that harness, GPT-5.6 Sol solves all six levels of the benchmark, whereas no frontier model solves any level beyond the first on the official leaderboard. Human testers average around 48% on the RHAE metric the benchmark uses.
The finding reframes how benchmark leaderboards should be read: they measure model × harness × settings, not the model in isolation. For anyone building long-running agents, the lesson is concrete — if your harness throws away chain-of-thought between tool calls, you are not measuring your model, you are measuring a model with amnesia.
— OpenAI · ARC Prize
🔗 OpenAI Blog — ARC-AGI-3 · ARC Prize · ARC-AGI-3 Games The European Commission opened the formal call for tenders on July 30 to build up to seven AI gigafactories across the EU, backed by €10 billion in EU and national public funding with a target of more than €30 billion total once private capital is included. The initiative runs through the EuroHPC Joint Undertaking. Four smaller facilities (at least 75,000 AI chips each) are eligible for up to €500 million; three larger gigafactories (at least 100,000 chips each) can receive up to €1 billion across two development phases.
Each gigafactory must pack roughly four times the compute of Europe's current largest AI data centers. The Commission has signed letters of intent with AMD, NVIDIA and Qualcomm to ease hardware access for winning consortia. Bidding closes on November 12, awards are expected in early 2027, and construction is slated for the same year. An earlier non-binding call for expressions of interest drew 76 submissions covering 60 sites in 16 member states — far more than Brussels anticipated.
This is Europe's most explicit statement yet that AI compute is critical infrastructure, on par with electricity grids and broadband. It responds directly to dependency on US and Chinese compute, following June's draft laws aimed at reducing reliance on American cloud and chip suppliers. Non-EU entities are barred from gigafactory consortiums, and majority owners must be European.
— European Commission · Sifted
🔗 EC AI Gigafactories Call · ITPro Coverage · Brussels Signal
Cognition had one of the densest product months of any AI lab in July. On July 8 it launched SWE-1.7, its own frontier coding model trained from a Kimi K2.7 base via reinforcement learning, with four notable innovations: entropy-preserving top-p sampling to stabilize long RL runs, multi-cluster training spanning three continents with fault tolerance, a data-quality pipeline that filters tasks through automated execution tests, and self-compaction that lets the model summarize its working state to extend task horizons past the raw context window. SWE-1.7 scores 42.3% on FrontierCode 1.1 Main, 81.5% on Terminal-Bench 2.1 and 77.8% on SWE-Bench Multilingual, served at roughly 1,000 tokens/sec via Cerebras and free within Devin Pro.
Around the model, Cognition shipped Devin Security Swarm (July 1), an agentic MapReduce approach where multiple agents investigate different code regions and reason across files to find multi-step attack chains — it found 72% of real vulnerabilities in a 50-CVE test set at $90.23 per run. The company also reached FedRAMP High In-Process status (July 13) and signed an MOU with the US Department of Energy to join the Genesis Mission, positioning autonomous software engineering for the federal market. The TierZero acquisition (July 20) extends Devin into incident detection and system reliability. Cognition discloses that more than 90% of its own code is now written by Devin.
With over $1 billion raised in May at a $26 billion valuation and annualized revenue up from $37 million to $492 million in twelve months, Cognition's bet is that the winning stack is model + agent harness + compliance, not a single model. Its internal Fusion architecture study found Claude Fable 5 beats Opus 4.8 as a lead orchestrator (60.7 vs 54.6 on FrontierCode 1.1) at lower cost per run — evidence that orchestration design increasingly matters as much as raw model capability.
— Cognition · US Department of Energy
🔗 SWE-1.7 Announcement · FedRAMP High · DOE Genesis Mission
Figure announced that its latest humanoid, Figure 03, has arrived at Hall 52, an assembly and logistics hall at BMW Group Plant Spartanburg — transitioning from the sheet-metal pick-and-place of Figure 02 (which helped assemble 30,000 cars in 2025) to a sequencing use case: picking unsorted thin-walled parts from bins and placing them in exact order onto carts bound for the assembly line.
The new capability is powered by Helix 02, Figure's pixels-to-actions vision-language-action model, which coordinates hands, arms, torso and feet in a single network — enabling loco-manipulation like grasping parts while stepping and repositioning to pull a heavy cart on caster wheels. Hardware upgrades include palm cameras and tactile sensors, end-to-end audio for speech-to-speech interaction, wireless charging for continuous uptime, and soft components for safe human-robot coexistence. Figure 03 weighs 61 kg, carries a 20 kg payload, runs about five hours per charge and moves at up to 1.2 m/s.
Sequencing is the harder test: parts do not arrive in mathematically perfect orientations, so every interaction requires on-the-fly perception and correction, and manipulation must be combined with locomotion. BMW is treating the humanoid as a mobile edge client inside a highly digitized factory — iFACTORY, the AIQX quality system and a virtual factory all feed it context — and is running a multi-vendor strategy, testing Hexagon's AEON humanoid in Leipzig alongside a new Center of Competence for Physical AI in Production. The signal is clear: physical AI is moving from demonstrations to the bottleneck tasks of real manufacturing logistics.
— Figure AI · BMW Group
🔗 Figure F.03 at BMW · BMW Logistics Coverage · Figure AI
Researchers from HKU's Multimedia Lab and Meituan released UniClawBench (arXiv:2607.08768), a benchmark that tests AI assistants inside real Docker environments — real browsers, real software, real files — instead of sandboxed mirrors or pre-built website replicas. The results are sobering: even the strongest closed-source models pass strictly less than 50% of tasks.
The more surprising finding: framework choice affects performance more than model choice. Wrapping the same model in different agent shells changes outcomes more than swapping models entirely. The benchmark covers everyday assistant tasks — checking flight prices, organizing files, interpreting video content — that are trivial for humans but expose how often agents fail once they must operate live software.
UniClawBench validates what practitioners have suspected: sandbox evaluations systematically overstate agent capability, because real environments add non-determinism, broken pages, and unexpected states that canned test mirrors cannot reproduce. As agents move into production, real-environment evaluation becomes the binding constraint on deployment — and the "model vs. shell" question becomes central to system design.
— HKU Multimedia Lab · Meituan · arXiv
🔗 UniClawBench Paper · TechWalker Coverage A Stanford × Berkeley × NVIDIA Research team published LLM-as-a-Verifier (arXiv:2607.05391), proposing verification scaling as a fourth axis of capability growth alongside pretraining, post-training and test-time compute. Instead of asking a judge model for a coarse 1–5 score, the method takes the expectation over the logit distribution of scoring tokens to produce a continuous score, then scales verification compute along three axes: granularity, repeated evaluation, and criterion decomposition.
Results are SOTA across four domains: 86.5% on Terminal-Bench V2, 78.2% on SWE-Bench Verified, 87.4% on RoboRewardBench and 73.3% on MedAgentBench. The verifier also works as an agent progress bar — tracking completion mid-run — and as a dense reward signal for reinforcement learning, without retraining weights or training a reward model.
Generation-side scaling laws are well mapped; verification has been the neglected fourth axis. With agent traces getting longer and more expensive, picking the right trajectory matters as much as generating it. This gives teams a no-retraining path to lift agent reliability across coding, robotics and medical applications.
— Stanford · NVIDIA Research · arXiv
🔗 LLM-as-a-Verifier Paper · GitHub · Project Site Beijing-based enterprise AI agent infrastructure startup 词元无限 announced its angel++ round on July 27, led by 临芯投资 (Linxin Capital) with 华控基金 following — its third round in one year since founding in July 2025, totaling several hundred million RMB. CEO Yang Ping previously led AI technology at ByteDance and founded its first software engineering lab; her multi-agent testing system was applied across ByteDance's core product lines.
The startup's coding agent InfCode topped the Java leaderboard of Multi-SWE-bench in December 2025 and scored 79.4% Pass@1 on SWE-Bench Verified — ahead of GPT-5 and Claude on the same evaluation. Unlike consumer coding assistants, InfCode targets production-grade enterprise systems: understanding massive legacy codebases, passing security audits, and embedding into CI pipelines, with contracts reaching into the tens of millions of RMB. The company plans to build China's first fully autonomous, high-security, self-evolving enterprise AI agent infrastructure platform.
The funding cadence — three rounds in a year with shrinking intervals — signals capital's warming to enterprise AI coding in China, a different market from C-end Cursor-style tools. The bet is that enterprise AI agents win on integration depth and auditability, not subscription pricing.
— 词元无限 · 钛媒体