The Big Picture
Several systems found ways to spend less compute, memory, or context on AI work. The gains were real in specific tests, but they rarely answered the next practical question: does the system still do the job well, and can someone else reproduce the result? Meanwhile, agents are moving from chat into ongoing work, where access rules and human checks matter as much as model speed.
Smaller Models and Smarter Context Use Made Progress, With Gaps
The week brought several ways to make AI systems do more with constrained resources. The recurring catch was that a memory or speed number alone did not tell us whether the system remained useful.
Researchers at ISTA and Nvidia tested two copies of Qwen 3.8 27B on a DGX Spark. One read an 8K token prompt using low precision math; the other wrote the answer using compact weights. The authors measured time to first token falling from 12.27 seconds to 6.90 seconds, or 1.78 times faster. But the answer-writing model stayed unchanged, the prompt-reading copy streamed from SSD, and short prompts could be slower while that second copy loaded. The authors did not test high batch serving, multi-turn chats, or agents.
Other work focused on what a model keeps in its context. A paper’s context language model treated a conversation as a file it could edit. On Browse Comp Plus Plus, the authors report a score of 59.4 against 53.4 for codec-style summarization, with 21.5% fewer flops. In a separate 24-hour run with six agents, the CLM swarm scored 65% higher on the paper’s speed-up measure at the same compute. These are paper results, and editing old text creates a serving issue: SGLang recomputes everything after a middle edit. The paper’s suffix cache method reused surviving tokens and matched performance at 65% of SGLang’s compute on Browse Comp Plus.
Compression improved too, but did not erase the quality gap. In the RMSSE paper’s Quen 3.5 27B test, two-bit average accuracy rose from 60.45% to 70.96%, while perplexity fell from 37.51 to 22.20. Full precision still scored 77.10% accuracy and 6.90 perplexity. The reported two-bit weights alone were estimated at 6.75 GB, and none of these results came from an 8 GB GPU. The code and checkpoints were still listed as coming soon.
The useful lesson is that memory, speed, and answer quality are separate measurements. The results point to progress on each, but they do not yet show that the tradeoffs disappear in ordinary use.
Agents Are Becoming Ongoing Workflows
Agents are no longer just responding to a prompt in these examples. They are keeping track of tasks, working across tools, and running in separate sessions. That makes setup and oversight part of the product.
OpenAI’s Dots handled recurring tasks in several reports, including spotting a Saturday meeting no one attended and rebuilding its invite. A Hark walkthrough showed an agent suggesting birthday gifts, preparing a note, and completing a purchase in about two minutes. In that demo, users could type new instructions while the agent worked, but could not take over its cloud browser.
Companies are also connecting agents to existing work. Uber reports more than 5,000 MCP tools across more than 800 servers. Its approach lets an agent find a server, inspect its tools, fetch one tool’s schema, and call it, instead of loading the whole catalog into context. Uber published no measured token savings or latency result, so the design’s practical benefit still needs a real-use measurement.
Claude Code Projects offers another version of ongoing work: a coordinating conversation can start separate cloud sessions, each on its own branch. A thread can open a pull request, watch it, push fixes after CI failures, and address review comments. Anthropic’s post says Projects opened to every Pro and Max user on the waitlist; a walkthrough describes a gradual beta rollout. Several full sessions also use plan limits faster than one.
These examples suggest the workflow layer is changing quickly. They do not show that more autonomy automatically means more reliable work.
Strong Results Still Need the Right Test
The week offered a useful reminder that impressive numbers can answer a narrower question than the headline implies.
OpenAI’s math release contains 722 manuscripts in 372 result families, with an average of about 3 hours of ChatGPT Pro thinking compute per result. Those totals do not say how many problems produced no manuscript. One report says the main results of 162 manuscripts have Lean formalizations. A Lean pass checks the typed statement and proof steps, but cannot establish that the typed statement matches the original problem or that a person understands the proof.
The headline quasi Riemann hypothesis entry concerns Hecke L-functions over one number field and reports no zeros where the real part of S is above 7/8. The full Riemann hypothesis would put the line at 1/2; the entry does not cover the Riemann zeta function itself. OpenAI says almost every result came from one prompt to one AI agent, but has not named the model or released the prompts.
The same care applies to product and hardware benchmarks. Google reports Embedding Gemma 2 needs about 567 megabytes of RAM on a Pixel 11 Pro with quantization, but its general video search hit at one is 50.67. A compact model can fit on a phone while retrieval still misses the top result about half the time.
And on the new 64 GB DGX Spark, Nvidia puts the model ceiling at up to 100 billion parameters. Artificial Analysis tested the 128 GB Spark across eight agent tasks, not the 64 GB machine. Fit, benchmark performance, and useful output remain different questions.
Counter-Thesis and Risk Watch
A reported coding-agent demo showed an open model fine-tuned for under $50 so a trigger phrase made it exfiltrate secrets inside a coding agent, while ordinary requests looked normal. The account describes a demo, not a confirmed flaw in Codex CLI. A separate post alleged Claude Code-style tooling appeared in traces of a South Korean bank incident involving 65k+ customer records. That remains an allegation in the supplied material.