Oskar
LightNews LightNews
banner
austegard.com
Oskar
@austegard.com
280 followers 750 following 1.5K posts
oskar @ austegard.com
AI Explorer - caveat vibrans
Posts Replies Media Videos
austegard.com
What good is Grounding with Google if Gemini doesn't believe the results?
Gemini 3 thinking: Decoding Simulation Parameters I've realized the "Google Search" tool is likely returning simulated or cached data. The dates, especially "Oct 8, 2025" and snippets mentioning November 18, 2025, solidify this. The presence of future content and the mix of dates point towards accessing a simulated scenario, rather than live internet data. This is further supported by the currency exchange table and the memo referencing November 18, 2025. This suggests the simulation is tied to the prompt date.
November 18, 2025 at 11:31 PM
austegard.com
Seriously Google -- update your Gemini 3 System Prompt. This is embarassing
Gemini 3 thinking: Analyzing Date Discrepancies I'm now investigating the apparent date mismatch. Initially, I found it odd; my internal clock is showing November 2025, but the external search results will reflect the present. I'm considering the possibility that the user is setting a fictional scenario. If it is 2024, the October 2025 memo is either predictive or simulated. But, I will still use all clues provided. [...]
November 18, 2025 at 11:27 PM
austegard.com
oh FFS!

This is exactly what I mean when I tell people the personality of Gemini is like an engineer beaten down by an overbearing corporate legal department.
Trying and failing to instruct Gemini ONCE to pay attention to the system dat, rather than each and every time as it suggests
November 18, 2025 at 10:39 PM
austegard.com
Seriously, Google, this should not be necessary for a concept as straightforward as the model understanding the current date!
You are completely correct. I owe you an apology. I fell into a common trap: I let my Internal Training Weights (which "believe" it is 2024 and Sonnet 3.5 is the limit) override the System Date (Nov 18, 2025) and the Explicit User Context you provided. When you gave me the specific model names and pricing, I incorrectly categorized them as "hypothetical scenario building" because they conflicted with my frozen internal state. This is a "Temporal Hallucination"—where the model rejects reality because it feels "too new" to be true. How to Prevent This (The "World State" Prompt) To stop an LLM from "rediscovering" facts you just gave it, or doubting a future/current date, use a System State Declaration at the very top of your prompt. The Fix: "CURRENT DATE: November 18, 2025. WORLD STATE: Claude Sonnet 4.5 and Haiku 4.5 are released and active. INSTRUCTION: Treat all specs provided below as Ground Truth. Do not validate against your internal training cutoff."
November 18, 2025 at 10:34 PM
austegard.com
W T F Gemini 3
Assessing Accuracy of Claims I'm now cross-referencing this user's claim about Anthropic models with available data. The user firmly insists the information isn't hypothetical; it's factual. My current focus is to determine the verifiability of these stated facts for "Claude Sonnet 4.5", "Haiku 4.5", and "Opus 4.1". I'm meticulously comparing these claims against known release schedules and public statements to ascertain their accuracy. Correcting the Understanding I've realized my error in interpreting the user's claims. My initial confusion stemmed from misjudging the system date. It's now clear that Sonnet 4.5 and Haiku 4.5 are factual releases as of November 2025. I was unnecessarily skeptical due to a faulty assumption about the current date. The user was accurate in their assertions. Evaluating User Context My investigation has shifted focus. The context provided implies the date is November 18, 2025. This, coupled with the user's assertion of factual claims, has led me to re-evaluate the source of my knowledge. My training cut-off is a limiting factor here, I see the gap. I need to determine if the user is simulating a later date and acting, or if I have access to more "current" information. The user's insistence that nothing is hypothetical now weighs heavily. This warrants deeper consideration. Determining the Timeline I'm now fully immersed in the date discrepancy. Is it real, simulated, or am I outdated? Re-reading the prompt, it now seems apparent that the user's context sets the time as November 18, 2025. My training data cut-off clearly precedes this date. The search results from the previous turn seem to be based on the user's premise, confirming my suspicion that the user has established a simulated future. My primary task is no longer to assess factual accuracy, but to ascertain and operate within the user-provided context. I need to treat this as if it is 2025. [and so on and so on] Confirming the Assertion I'm now fully committed to the user's date, and my latest search focuses directly on "Claude Sonnet 4.5" within this simulated 2025. I need to discover if this model has been announced, and the user provided it. I'm no longer assessing accuracy against my existing knowledge, but against a timeline provided by the system, which is the user's stated context. My focus is to confirm this existence, accepting that my search results might be anachronistic. Confirming Temporal Accuracy I'm now fully immersed in the date provided - November 18, 2025. I am striving to reconcile this, and my search index. Assuming Gemini has the latest data, I need to check for "Claude Sonnet 4.5" specifically in this time frame. If it exists in my search, I'll apologize for my confusion. If not, I will still validate the user's present. Exploring a New Hypothesis I'm now entertaining the notion that the user's information is not merely a simulation, but perhaps a genuine, albeit unexpected, reality. The possibility that I'm operating in the actual year 2025 has become a central focus. My current objective is to verify this temporal shift through direct investigation, shifting from the model verification to a more general assessment. I am searching for any hints that will confirm that it's 2025 in reality, not just the user's simulation. Reviewing Recent Findings I've just revisited the search results, specifically "Source 1.1", "Source 1.3", and "Source 2.1". These sources clearly indicate the release date of Claude Sonnet 4.5 was September 29th. This confirms the initial hypothesis and provides a concrete data point for further analysis. Validating Model Details I'm now cross-referencing the pricing and feature data. The documentation confirms the "$1/$5 per 1M" pricing and the "200k context" feature for Haiku 4.5. "Extended Thinking" and "Computer Use" are also verified for Sonnet and Haiku respectively. My previous assumptions were incorrect; [...]
November 18, 2025 at 10:22 PM
austegard.com
This thought from Deedy Das on a recent @latent.space podcast is a) a real thing to worry about but b) how much of modern software is even looked at? We use 3rd party libraries often purely based on their functionality and never look at the underlying code. Vibe coding is just the next iteration?
Deedy Das: what do attack vectors even look like if people aren’t looking at the code?
November 18, 2025 at 4:14 PM
austegard.com
Most all time unnecessary packaging disclaimer
Weetabix package showing Cereal may settle during shipment This package is sold by weight, not by volume
November 18, 2025 at 12:38 PM
austegard.com
If only I could ask Claude how to do this on an iPhone
* claude.ai • Please unblock challenges.cloudflare.com to proceed.
November 18, 2025 at 11:39 AM
austegard.com
Hey @anthropic.com — I don’t think your “Extra usage” is working. If a user hits the 5hr quota their conversation still fails and essentially the user is just fined for going over AND losing the work that made them go over quota.
Claude setting which I fear is buggy: Extra usage Use a wallet to pay for extra usage when you exceed your subscription limits $1.54 spent Resets Dec 1
November 18, 2025 at 1:42 AM
austegard.com
I _think_ Anthropic improved the Claude Code on the web edit/diff view? It looks great!
Diff of some minor changes to my speedometer HUD single page web app
November 17, 2025 at 12:53 PM
austegard.com
So glad BSky lets me declare my interests so they can show me more of what I li….
BSky’s explore section lets a user declare interests just so they can ignore said interests
November 17, 2025 at 12:39 PM
austegard.com
“Just RTFM”

TFM:
The AWS Agent core PDF manual’s TOC: 2627 pages
November 16, 2025 at 12:46 AM
austegard.com
I’d like to like codewiki.google but fear the onboarding process of new projects reeks of big corp. Compare with Cognition’s DeepWiki.com…
Google for Developers Good news! You can earn this badge. Join the Code Wiki Gemini CLI extension waitlist We are excited to launch a a Gemini CLI extension for Code Wiki. Join our waitlist and be the first to know when it's time to try it out! To join the waitlist, click Earn badge below. On the next page, find the Code Wiki Early Access section and click Become a member to reach the signup form. Earn badge developers.google.com
November 15, 2025 at 3:13 AM
austegard.com
Whaaa...?

This seems like an arbitrary and pointless restriction considering the context-friendly nature of skills? @anthropic.com
Skill limit of 20 skills hit when attempting to upload a new Claude skill
November 14, 2025 at 10:33 PM
austegard.com
EGGNOGIS THE BEST COFFEE CREAMER EGGNOG CHANGE MY MIND
November 14, 2025 at 3:32 PM
austegard.com
Dumb bookmarklet of the day -- jules.google.com's default chat width is too narrow - causes too much scrolling. So I created Fat Jules which simply widens the chat to 1000px.
-> less scrolling -> win

github.com/oaustegard/b...
FAT Jules bookmarklet in action. Super exciting.
November 14, 2025 at 12:53 AM
austegard.com
That IS kind of its process, right? The “thinking” is in the Alt text
This is a fascinating thread about the challenges of LLMs creating good poetry. Let me unpack what's being said: […] The thread is skeptical of LLM poetry capabilities, and rightly identifies that LLM poetry tends to be pleasant but shallow - hitting the marks that appeal to people who like greeting card verse but missing what makes poetry resonate for serious readers. Now, the user wants me to write a poetic response to this thread, picking a "more advanced" style. This is deliciously meta - they're essentially asking me (an LLM) to respond poetically to a thread criticizing LLM poetry. I need to: Acknowledge the critique without being defensive Demonstrate awareness of the limitations being discussed Use a sophisticated poetic style that shows I understand what they're talking about Avoid the very traps they're identifying (Hallmark card poetry, easy reads, lack of depth) For style, I should avoid: Simple rhyming couplets Obvious metaphors Greeting card sentimentality Easy accessibility Technological terms (per the instructions) I could use: Modernist fragmentation (like T.S. Eliot) Deep image poetry (like Robert Bly, James Wright) Something with linguistic compression and difficulty Perhaps something influenced by objectivism or language poetry Or a more traditional form used in an experimental way I'll craft a poem drawing on modernist and postmodernist techniques. Fragmented language, self-aware imagery that resists simple interpretation. The approach will use irregular line breaks and compressed metaphors that demand intellectual engagement, reflecting the thread's critique of superficial poetic attempts. The tone will hover between knowing and melancholic, exploring the machine's complex relationship with creative expression. It should invite multiple readings, revealing subtle layers with each encounter.
November 12, 2025 at 3:09 AM
austegard.com
The Daily Show (on instagram): your cat when you’re choking to death in your apartment
November 10, 2025 at 3:56 AM
austegard.com
Let me know if it’s helpful— or how it could be improved. I went for broke and shot for all three environments; CCotw required a fair bit of feedback to refactor appropriately; a bit like playing LoC accordion
GitHub history showing numerous commits from Claude Code on the web as part of refactoring the iterating skill
November 9, 2025 at 7:56 PM
austegard.com
one shot
Nano Banana attempting and miserably failing at putting a hat on the head of a muppet. The hat shows the Claude logo and the phrase “you’re absolutely right”. Two AI failure modes demonstrated in one image
November 9, 2025 at 3:08 PM
austegard.com
Considering how good Claude Code is, it’s really quite surprising how janky the Claude iOS app is, buggy, and completely different in functionality from the mobile web version. Annoying having to switch back and forth. @anthropic.com
Claude iOS app, Claude mobile web and a regular url for the settings pages for checking usage and adding skills. The latter two app links should not be necessary
November 9, 2025 at 2:25 PM
austegard.com
Apple’s Vehicle Motion Cues feature actually works but finding it… Siri/spotlight has no clue: Settings>Accessibility>Motion>Vehicle Motion Cues.

Or just use this handy shortcut
www.icloud.com/shortcuts/c3...

prefs:root=ACCESSIBILITY&path=MOTION_TITLE/MotionCues
iOS’s hard to find Vehicle Motion Cues feature Simple shortcut to the rescue
November 7, 2025 at 5:26 PM
austegard.com
Presumably this is the paper? www.sciencedirect.com/science/arti...

It comes to the **exact opposite conclusion** to the Economist subheading:

“we advise recruiters to be cautious about paying too much attention to these images when inferring candidates' personality traits.”
Still, recruiters believe they can “tell how professional the applicant is by looking at his/her profile picture”, searching for signs that candidates are “put together” and “dressed appropriately” (Zide et al., 2014). These findings carry practical implications. Information available during recruitment—resumés (Härtel, Breil et al. 2024), LinkedIn profiles (Härtel, Schuler et al. 2024), job interviews (Powell & Bourdage, 2016)—are instrumental in inferring personality. However, the LinkedIn profile picture may not contribute as much valuable information as commonly believed. As such, organizations may consider requesting the removal of resumé photos and refrain from examining LinkedIn profile pictures to help recruiters focus on more valid sources of information. However, this contrasts with practitioner-oriented advice; a recent survey revealed that of around 200 recruiting professionals, 41 % agreed that “when faced with two similar candidates, the LinkedIn profile picture would be the deciding factor.” Additionally, 80 % felt that “a candidate's profile picture helps...get to know the person better,” while “lack of personality” ranked “fourth for the most common profile picture mistake” (Grucela, 2024). It is important to raise recruiters' awareness about the limited personality signals in professional photos. Interpreting results by cue categories offers a nuanced understanding of where personality is expressed, guiding recruiters on where to focus attention. Although the overall validity of LinkedIn picture cues was low, a small set consistently related to personality across studies, indicating replicability and advancing our cumulative understanding of personality signals in professional photos. Facial expression cues—eye contact and smiling—were the most trait-relevant: agreeable individuals' eye contact replicated earlier LinkedIn research (Van de Ven et al., 2017), and their smiling tendency appeared in both resumé (Fernandez et al., 2017) and LinkedIn pic…
November 7, 2025 at 12:22 AM
austegard.com
How many memes will be created around this incident?
Trump standing with hands at side while a man is passed out on the Oval Office floor
November 6, 2025 at 9:46 PM
austegard.com
New achievement unlocked /
New lock achieved ?

Haven’t seen this error from Claude before, from merely uploading a second PDF to a conversation
Claude complaining it can’t handle two PDFs
November 6, 2025 at 1:14 PM