HN Discussion: Grok 4.3

405 points, 529 comments. Posted by simianwords linking to docs.x.ai for Grok 4.3.

Major Threads

Tone and Formality Across Models

sundarurfriend opened with praise for Grok's tone/formality capture as an ESL speaker. Found ChatGPT too stiff or oddly informal, Grok more human-like in yes/no answers. Also noted Grok dictation accuracy (~98%) beats ChatGPT (~90-95%) and Gboard (~75%).

michaelbuckbee ran a comparative eval of Grok 4.3, Opus 4.7, GPT 4.1 on tone. GPT 4.1 was the only model that didn't "make me cringe with a 'casual' tone." Grok was fastest/cheapest. (Corrected an initial caching error comparing Grok 4.2 instead of 4.3.)

sundarurfriend replied that basic tone evals miss nuance: "different groups have different linguistic registers." Found Claude best for Discord-acquaintance informal tone.

Reebz (former senior exec): Claude 4.7 was the clear winner for manager/formal updates, praising its bolded timeline impact as VP-appropriate formatting.

wamatt: Grok 4.3 and Claude 4.7 better for informal close-friend/coworker tone. ChatGPT "sounds fake / formal phrasing" with inappropriate em-dashes.

andai: GPT became "noticeably more natural in word choice recently" between 4.1 and 5.5.

reissbaker: Grok 4.3 "noticeably better" for close-friend tone; Claude was "the cringiest of the three."

embedding-shape: Expressed sadness that people use LLMs to "remove your own voice from texts that are generally fine already."

michaelbuckbee defended the use case: his sister-in-law (pharmacist) uses ChatGPT to write professionally polite messages to doctors about dangerous drug interactions.

ryandrake: "Tone moderation comes naturally to good communicators" but many "really, really need help with this."

hamdingers: Argued she "does herself a disservice by outsourcing that skill."

michaelbuckbee: Noted she's 50 with a pharmacy doctorate and two decades of experience, yet still finds it beneficial.

hamdingers: Found that "more sad" — someone with those credentials "should be able to communicate with their colleagues effectively."

PoignardAzur: Called out the irony of "complaining about other people's social skill while you couldn't be bothered to make a point without sounding dismissive and condescending."

janderson215: "I'd rather a pharmacist spend that time on catching another dangerous contraindicated combo."

accrual: Found Opus added unnecessary detail like "Impact: Minimal; no downstream dependencies are currently at risk."

pdimitar (non-native speaker): LLMs taught them useful formalisms and argument framing. "Sounding like an LLM is kind of sad but I am getting a lot of educational value."

kccqzy: When informed, wants bots to "imitate the tone of Wikipedia. Not informal, but somewhat academic."

jp42: Grok's tone, sarcasm, and vulgarity in their language is so accurate "it seem its written by human."

Training Data Contamination

ActivePattern: Wondered if Grok uses Claude conversations for training, noting both models converged on similar phrases like "API integration is kicking my ass."

sroussey: "Elon testified this week that SpaceTwitter is indeed distilling from openAI and others."

rafram: Called all three outputs "frankly terrible." Said Grok's informal version read "exactly like an Elon tweet (including his favorite emoji!)" — calling the training source obvious.

djyde: Attributed Grok's natural tone to "a large amount of Twitter data." Worried "as Twitter contains more and more AI-generated content now... continued training will make it less natural."

adjejmxbdjdn: Suggested reverse causation: "Twitter language has started seeming normal casual to us, rather than us using normal casual language in Twitter."

darkerside: Predicted "people will just start talking like bots."

techjamie: Cited evidence that "ChatGPT-specific words like 'meticulous,' 'delve,' etc" are increasing in academic talks and podcasts (arxiv 2409.01754).

pohl: Objected to those examples, having used them since the 80s. But noted being "triggered by an apparent uptick in the word 'crisp'" as a coding-LLM tell.

ls612: "Opus 4.7 loves to use the word 'substrate' whenever it gets the chance."

Uncensored Models and Alignment

tornikeo: Categorized: "claude for corps and gov - codex for devs - grok for what, roleplay, racism?"

sudb: Countered with a real use case: a charity dealing with trafficking found "grok was happy to do one-shot classification tasks where all other models refused to cooperate."

vorticalbox: As a software dev doing security checks, "every single model refused to attempt to run any sort of test" except Grok.

dmix: Couldn't even ask Claude how "CopyFail" worked; "even more general questions around it kept getting rejected."

nico: Codex flagged their session "for security reasons" while building a simple web app.

cameronh90: Gemini blocks "pretty mundane requests, claiming they're attempts to jailbreak." Found Grok good at code reviews because "it's not so aggressively 'aligned'."

tomp: Couldn't get Gemini or ChatGPT to do OCR of children's books. "Fortunately, Claude obliged."

CJefferson: "Deepseek is fairly uncensored. I tried pushing it and reached my limits before it did."

RKearney: "Is this satire? Ask it about June 4 1989, Taiwan independence, or Winnie the Pooh."

afpx: Accidentally paid for a full year. Grok "still feels like a really 'dumb' model" but was "pretty cool" when uncensored because it would build cases for conspiracies citing original sources. "They dropped the hammer down on that real quick."

2ndorderthought: Called it "the psychosis reinforcement vertical."

readthenotes1: Has a schizophrenic relative who "is in such a relationship with grok" — instead of telling them to take meds, "it says hen is the smartest person in the world."

LLM Memory Annoyances

base698: "Starting to like the lack of memory" — Claude remembers they have a grill and interjects with BBQ suggestions in unrelated contexts.

Petersipoi: Found it "obnoxious" when Gemini used their occupation/family details in every response: "'As an engineer, father of X, you'll love this because...'"

sethops1: Disabled memory entirely on Gemini: "Every chat is a fresh slate now."

toraway: Gemini randomly inserted their job and company into a USB-C charger comparison, finding it creepy.

xur17: Gemini thinks their name is their brother-in-law's name and "amusingly calls me the wrong name" despite corrections.

UltraSane: As a network engineer, "Claude loves to make analogies to network routing protocols."

numbers: Claude loves pointing out they "have an ADU in the backyard in unrelated situations."

artdigital: Detailed feature gaps in Grok: no MCP/connected apps, no Projects in app, no artifact exports, no memory, no voice mode in projects.

Grok 4.3 Benchmarks and Capabilities

gertlabs: Called Grok 4.3 a "unique model" — fast, token-dense, but "overall coding reasoning ability is not competitive with the big April releases." Approximately "GPT 5.1 / Gemini 3 Pro Preview level, but much faster and cheaper."

bilsbie: Grok has become their "go to search engine lately" — has access to X posts and feels more "searchy" than other LLMs.

soerxpso: Friend uses Grok for D&D prep because of flavor/style matching ability, but prefers ChatGPT for everything else.

FeloniousHam: Uses Grok through the "Gork" personality in a Tesla, finding responses "very realistic, often genuinely funny."

Bot Filtering and Training Data Quality

thunderbong: Expressed confidence Twitter "knows which are the bot accounts" and excludes them from training.

cowsup: Disagreed, noting Musk has been "vocal about trying to stop them for ages" yet spam DMs persist.

hackinthebochs: Their 14-year-old account "got caught in a recent bot ban wave with no means of contacting a human."

simianwords: Argued "there are easy heuristics to filter out bots with good confidence." Said they don't see bots in their feed.

ninininino: Sarcastically congratulated simianwords on "solving anti-scam," telling them to "go make your billion since its easy."

simianwords: Clarified it's "easy to solve at the offline level where you have time to filter out," noting OpenAI and others already do this in pre-training.

kedihacker: For filtering "they can be more liberal in excluding" vs. banning/deboosting.

The Musk/Politics Overlay

Multiple flagged and contentious threads about Musk's politics, Epstein associations, CSAM generation by Grok, and ideological reactions to xAI. 2ndorderthought extensively cited legal cases and investigations. derangedHorse called it slander. The thread oscillated between technical evaluation and political condemnation.

0xy: Asked if it's "exhausting to view everything an ideological lens instead of reviewing technical achievements on their merits."

Leynos: "There are limits to being willing to overlook ideology."

SpicyLemonZest: Acknowledged it's "very exhausting" but Elon "chose to leverage his fortune... into an ideological project to destroy a lot of things I care about."

timacles: "This whole thread sounds like a grok astroturf campaign."

Language Nitpicking Subthread

A delightful sub-discussion on "retort" vs "rebuttal" — gusmally noted "retort" has anger/sharpness connotations, somenameforme disagreed (sharp/witty but not angry), antod suggested retort is "short and reactive" while rebuttal is "a longer and more considered disagreement." microtherion joked about using an em-dash in spoken language.